feat: add aimlapi.com as an OpenAI-compatible LLM provider

[Lookoff-AIMLAPI]

What

Adds aimlapi as an API provider: one OpenAI-compatible endpoint
(https://api.aimlapi.com/v1) in front of 350+ chat models from many vendors,
so a user can move between Deepseek-, Anthropic- and OpenAI-class models by
editing provider_model alone.

Shape copied from the existing gateway/API providers (openrouter_fn,
minimax_fn, litellm_fn — the litellm PR is the closest template):

• sources/llm_provider.py — aimlapi_fn, one entry in available_providers,
  one entry in unsafe_providers. The env var name falls out of the existing
  get_api_key() convention as AIMLAPI_API_KEY, no extra code.
• tests/test_aimlapi_provider.py — 16 tests, matching the one-module-per-provider
  convention of test_litellm_provider.py / test_minimax_provider.py.
• .env.example, README.md — key placeholder and a provider table row.

Base URL is overridable with AIMLAPI_BASE_URL, same as MiniMax.

Privacy

docs/CONTRIBUTING.md requires that cloud services be "optional alternatives,
clearly defined with a warning message". aimlapi is therefore in
unsafe_providers, so the existing "Warning: you are using an API provider.
You data will be sent to the cloud." fires, and is_local = True is refused.
Nothing about the default local path changes.

Two things that would have been bugs

1. Unset parameters are omitted, never sent as None. aimlapi.com answers
   400 for an explicit null on temperature, top_p, seed, tools,
   tool_choice, response_format, stream and max_tokens — but only on
   some models, and the OpenAI SDK serialises an unset optional as literal
   null. A provider copied verbatim from a working OpenAI-compatible one can
   therefore fail on every real call with a fully green mocked test suite.
   There is a regression test asserting no request parameter is None.
2. Attribution headers are keyed on the request origin, not the provider
   name. If AIMLAPI_BASE_URL is pointed at a proxy or a self-hosted gateway,
   aimlapi_attribution_headers() returns {} so nothing rides to a third
   party, and it returns a fresh dict every call so the module-level constant
   cannot be mutated. Both are covered by tests, as is the
   ^part_[A-Za-z0-9]{1,64}$ shape of the partner id — a malformed id is
   dropped silently at the gateway, so only a test catches a typo.

Verification

Environment: macOS arm64, Python 3.12 venv from requirements.txt. vosk==0.3.45
publishes no macOS arm64 wheel, so it was omitted from the local install; it is
unrelated to this change and nothing in the test suite imports it.

Tests — python -m pytest tests/ -q

	passed	failed	subtests
baseline (main, pristine)	154	9	12
with this branch	170	9	12

The 9 failures are identical in both runs and pre-existing: 8 in
tests/test_litellm_provider.py (litellm's current API surface) and
tests/test_memory.py::test_save_and_load_memory. The 16 added passes are the
new provider module.

Live inference through Provider("aimlapi", ...).respond() — not a mock,
not a raw curl. Both models shipped in the README row were called:

catalog check deepseek/deepseek-v4-flash: id=True alias=False
catalog check anthropic/claude-sonnet-4.6: id=True alias=False

=== deepseek/deepseek-v4-flash ===
base_url        : https://api.aimlapi.com/v1
default_headers : {
  "HTTP-Referer": "https://github.com/Fosowl/agenticSeek",
  "X-AIMLAPI-Partner-ID": "part_agenticseek",
  "X-AIMLAPI-Source": "agent/agenticseek",
  "X-Title": "AgenticSeek"
}
response        : 'AIMLAPI OK'

=== anthropic/claude-sonnet-4.6 ===
base_url        : https://api.aimlapi.com/v1
default_headers : { ...same four headers... }
response        : 'AIMLAPI OK'

Model ids were checked against both id and aliases in
GET /v1/models?include=all filtered to type == "openai/chat-completions",
and then actually called — that catalog both omits ids that work and lists
at least one that does not, so membership alone is not proof.

Two-turn tool loop — CoderAgent driven end to end on
deepseek/deepseek-v4-flash: turn 1 the model emits a bash block, the framework
executes it and pushes the output back as a user message, turn 2 the model
answers from that output. Final memory:

[0] system: ...coder_agent.txt...
[1] user:   Using a single bash block, run exactly `echo AIMLAPI_TOOL_TURN_1`. ...
[2] assistant: ```bash / echo AIMLAPI_TOOL_TURN_1 / ```
[3] user:   [success] Execution success, code output: ... AIMLAPI_TOOL_TURN_1
[4] assistant: It printed AIMLAPI_TOOL_TURN_1.

Final answer: It printed AIMLAPI_TOOL_TURN_1.

Not verified / caveats

• max_tokens does not reliably bound reasoning tokens on this API — some
  reasoning models return several times the cap with finish_reason: "stop".
  AgenticSeek exposes no max-tokens control, so nothing here depends on it, but
  it is worth knowing before one is added.
• Only README.md was updated. The 7 translated READMEs and the stale
  "API Providers (Cloud-Based)" table further down README.md were left alone —
  neither lists openrouter, minimax or litellm either, and PR Upgrade MiniMax provider: add M3 as default, drop M2.5 variants Fosowl/agenticSeek#511
  (MiniMax) touched only README.md + README_CHS.md, so full fan-out is not
  the convention here.
• vosk could not be installed locally (no macOS arm64 wheel), so any test that
  would need it was never exercised — none appear to exist.

Commits

• feat: add aimlapi.com as an OpenAI-compatible LLM provider
• chore(aimlapi): fork-only placement — do not send upstream — moves the
  provider row to the top of the hand-ordered README table. Drop this commit
  before any upstream PR; the provider is complete without it. No badge was
  invented: this repo has no featured/recommended concept.