fix: normalize local tool JSON fallbacks

[scubamount]

Summary

• normalize local-model assistant JSON with commands/tools_used into OpenAI-style tool calls
• support direct argument fields as well as nested arguments objects
• allow execute_code(command=...) compatibility by wrapping shell commands into Python subprocess code while preserving code= precedence

Tests

• PYTHONPATH=/Users/andrew/.hermes/hermes-agent /opt/homebrew/opt/python@3.12/libexec/bin/python3 -m pytest -o addopts='' tests/agent/transports/test_chat_completions.py tests/tools/test_code_execution.py -q
• 141 passed in 98.56s

[scubamount]

Cross-ref: this PR is the client-side counterpart to jundot/omlx#1596 (feat: add Hermes tool call parser).

These two PRs are co-dependent:

• hermes-agent#37229 (this PR): normalizes local-model assistant JSON with commands/tools_used into OpenAI-style tool calls on the hermes-agent side.
• omlx#1596: parses Hermes-style <|tool_call_start|>...<|tool_call_end|> markers at the OMLX inference server layer, returning structured tool calls instead of raw marker text.

Without both, local model tool calling is broken:

• Without hermes-agent#37229 → hermes-agent cannot normalize mixed JSON/bracket payloads from local models
• Without omlx#1596 → OMLX leaks raw Hermes markers that hermes-agent cannot parse

Both should merge together for a complete fix.