feat(anthropic): prompt caching with cache_control: ephemeral

[fede-kamel]

Closes #35.

Problem

Locus's `AnthropicModel` didn't expose Anthropic's prompt caching. Long system prompts and large tool catalogs paid full input-token cost on every turn — even when 4-5k tokens of stable content could be cached at ~1/10× cost for ~5 minutes.

Fix

Opt-in via `AnthropicModel(prompt_cache=True)`. Default `False` for backward compatibility.

When enabled:

1. System prompt is sent as a block list with `cache_control: ephemeral`:
   ```python
   params["system"] = [
   {"type": "text", "text": system_prompt,
   "cache_control": {"type": "ephemeral"}}
   ]
   ```
2. Tool catalog — the last tool entry carries the same `cache_control` marker (Anthropic walks markers in order; the last one anchors the cache point, covering everything before it).
3. Cache token counts flow through:
   • `response.usage.cache_creation_input_tokens` / `cache_read_input_tokens` extracted in the Anthropic provider
   • Surfaced on `ModelResponse.usage`
   • Accumulated in `AgentState` via new `with_token_usage(..., cache_creation_tokens=, cache_read_tokens=)` kwargs
   • Surfaced on `ExecutionMetrics.cache_creation_input_tokens` and `.cache_read_input_tokens`

Backward compatibility

• `prompt_cache=False` (default) preserves the bare-string `system` and untagged tools — zero behaviour change for existing callers.
• Old Anthropic SDK responses without the cache fields work unchanged — the keys just don't appear in `usage`.
• `AgentState.with_token_usage()` cache kwargs default to 0 — non-Anthropic providers see no change.

Tests

Seven new unit tests in `tests/unit/test_anthropic_prompt_caching.py` mock the Anthropic client and verify each layer of the wiring. All 3,212 pre-existing unit tests still pass; pre-commit clean; `mkdocs build --strict` clean.

How to use

```python
from locus import Agent
from locus.models.native.anthropic import AnthropicModel

agent = Agent(
model=AnthropicModel(
model="claude-sonnet-4-20250514",
prompt_cache=True, # turn it on
),
tools=[...],
system_prompt="",
)

result = agent.run_sync("...")
print(f"cache writes: {result.metrics.cache_creation_input_tokens}")
print(f"cache reads: {result.metrics.cache_read_input_tokens}")
```

Test plan

• `hatch run test tests/unit/` — 3,212 pass + 1 skip (OCIOpenAIModel needs real config)
• `pre-commit run --all-files` — clean
• `hatch -e docs run mkdocs build --strict` — clean
• DCO sign-off