feat(sdk): LlamaIndex callback adapter

[prashar32]

Adds a LlamaIndex integration to the Python SDK so a LlamaIndex query or agent is governed by a RiskKernel run with ~no code change — the LlamaIndex analog of the existing LangChain adapter.

What it does

RiskKernelCallbackHandler (in riskkernel/adapters/llama_index.py) is a LlamaIndex BaseCallbackHandler:

• Ticks one governed step per LLM call (CBEventType.LLM on_event_start), so a run's loop/time budget is enforced over the outer LlamaIndex loop. The next step over budget raises BudgetExceeded, halting the query/agent.
• The halt surfaces, not swallowed: LlamaIndex's CallbackManager.on_event_start doesn't wrap handler calls in a try/except, so the exception propagates to the caller — no extra flag needed (unlike LangChain's raise_error). I verified this in the LlamaIndex source and with a real integration test.
• Optional tool gating (gate_tools=True) routes CBEventType.FUNCTION_CALL events through the approval gate, mirroring the LangChain adapter's gate_tools option.
• llama-index-core is imported lazily with an object fallback base class, so the module imports and the SDK installs without llama-index present — it is not a dependency of the SDK. Pinned to the llama-index-core >= 0.10 callback protocol (documented in the module docstring).

Register it on Settings.callback_manager and governance is invisible until the budget bites:

from llama_index.core import Settings
from llama_index.core.callbacks import CallbackManager
from riskkernel.adapters.llama_index import RiskKernelCallbackHandler

Settings.callback_manager = CallbackManager([RiskKernelCallbackHandler(run)])

Callback events hooked

• CBEventType.LLM ("llm") — on_event_start ticks run.step() (loop/time enforcement).
• CBEventType.FUNCTION_CALL ("function_call") — on_event_start gated through the approval gate when gate_tools=True.
• on_event_end / start_trace / end_trace — implemented as no-ops (the protocol requires them); on_event_start returns the event_id so the callback manager can correlate the matching end event.

Tests

tests/test_llama_index.py exercises the governance path on stdlib alone against a fake Run: per-LLM-call step tick, loop-budget halt propagation, non-LLM events not ticking a step, tool-gating allow/deny, and event-type (CBEventType-like) normalization. A skip-unless-installed integration test proves a runaway loop of real LlamaIndex LLM calls actually stops at the budget. I ran the integration test against llama-index-core 0.14.22 in a throwaway venv and it passes; the full suite passes stdlib-only (25 tests, 3 framework-gated skips):

python -m unittest discover -s tests -t . -v
Ran 25 tests in 0.545s
OK (skipped=3)

README and CHANGELOG updated.

Closes #86.