fix(python): dispatch vindex infer through inference weights

[citizenu03bb]

Title:
fix(python): dispatch vindex infer through inference weights

Summary:

• Make Python Vindex lazy-load InferenceWeights instead of dense-only mmap weights for infer.
• Route infer and infer_trace through InferenceWeights::infer_patched.
• Keep dense-only analysis helpers using InferenceWeights::as_weights().

Why:
The Python Vindex.infer path used the dense mmap loader directly. That loader
does not understand Q4K/kquant vindex weight files, so quantized indexes could
load through the wrong path and produce bad predictions/traces. The shared
InferenceWeights loader already detects the format and dispatches to dense or
Q4K inference as appropriate.

Verification:

• cargo check -p larql-python
• cargo test -p larql-python --lib

Branch:
pr/python-kquant-infer-dispatch

Commit:
e4fe4fd