perf: raise ONNX default batch size to 64

[j-sperling]

Summary

• Raises OnnxEmbedding._DEFAULT_BATCH_SIZE from 32 to 64. Measured on the default gpahal/bge-m3-onnx-int8 (CPU, Apple M-series, 128 texts x ~230 tokens): 14.1s @ batch 16, 13.2s @ 32, 11.7s @ 64, 10.6s @ 128 — so 64 is ~11% faster than the current default. On corpus-scale indexing (observed >4.5h for 186K chunks with this model) that's tens of minutes.
• 128 is deliberately not the default: the tokenizer pads to the longest text in a batch with max_length=8192, so a worst-case batch of long texts at 128-wide materializes multi-GB activation tensors. 64 keeps the worst case bounded while capturing most of the win; users can still set embedding.batch_size = 128 in config where memory allows.
• embedding.batch_size config continues to override; 0 still means "provider default".

Test plan

• Full suite passes (no test pins the ONNX default batch size)
• ruff check / ruff format --check clean
• Timing measurements above reproduce with a 4-text warmup + time.perf_counter around embed()