feat: optional CUDA/GPU acceleration for embeddings

[sattva1]

Summary

Closes #64.

Adds opt-in CUDA support for the embedding pipeline, reducing embedding time from minutes to seconds on GPU-capable machines.

Two activation paths:

• --cuda CLI flag on analyze, watch, host, and serve
• AXON_CUDA=1 environment variable (works across all commands without per-command flags)

Design

Uses a module-level configuration in embedder.py rather than threading a cuda parameter through every function signature. _get_model() reads the CUDA state at call time via _resolve_cuda(), which checks both the programmatic flag and the env var. This means all embedding call sites — pipeline, watcher, and search-time embed_query() from MCP/web — automatically use GPU when enabled.

CUDA validation

When CUDA is requested, _get_model() captures fastembed's RuntimeWarning on CUDAExecutionProvider fallback and converts it to a RuntimeError with actionable install instructions. CLI commands call validate_cuda() before the pipeline starts, so the error surfaces immediately rather than being swallowed by the pipeline's broad except Exception handler.

Changes

• src/axon/core/embeddings/embedder.py — configure_cuda(), _resolve_cuda(), validate_cuda(). _get_model() uses (model_name, cuda) compound cache key and post-init CUDA fallback detection.
• src/axon/cli/main.py — _configure_and_validate_cuda() helper. --cuda flag on analyze, watch, host, serve.
• tests/core/test_embedder.py — 15 new tests covering flag, env var, cache separation, fallback detection.

Usage

# Via CLI flag
axon analyze --cuda .

# Via environment variable (works with any command)
export AXON_CUDA=1
axon analyze .
axon watch
axon host

[sattva1]

Two additional fixes pushed:

1. Explicit cuda=False in CPU path (2365c99)

fastembed defaults to Device.AUTO, which auto-detects and uses CUDA when onnxruntime-gpu is installed — even without --cuda or AXON_CUDA. This caused silent OOM on GPUs with limited VRAM. The CPU path now explicitly passes cuda=False to override auto-detection.

2. Batch size 32 → 8 (acac2e8)

The default batch size was never adjusted when the model changed from BAAI/bge-small-en-v1.5 (6 heads, 512-token context) to nomic-embed-text-v1.5 (12 heads, 2048-token context). Each batch element's attention matrix is ~192 MB with nomic — at batch_size=32 that's ~6.4 GB for attention alone, causing OOM on both CPU and 8 GB GPUs. Batch size 8 keeps peak attention memory under ~2 GB.