Fix local embedding context rebuilds and batch safety

[andinux]

Summary

Fixes local embedding stability and context handling for sqlmem / sqlite-memory.

Changes

• Skip whitespace-only semantic chunks before embedding.
• Size local llama contexts from max_tokens + overlay_tokens.
• Configure n_ctx, n_batch, n_ubatch, and token buffers consistently to avoid llama.cpp encoder assertions.
• Use reusable llama batches with explicit sequence metadata.
• Truncate over-capacity tokenization safely instead of relying on llama.cpp failures.
• Move llama diagnostics out of process-global logger user_data into thread-local storage.
• Rebuild the local engine when max_tokens or overlay_tokens changes after memory_set_model.
• Roll back option updates if local engine rebuild fails.
• Clear cached embeddings for the active local provider/model after context rebuilds.
• Add regression coverage for whitespace-only chunks and stale logger user_data.
• Bump extension version to 1.2.1.

Why

Local embedding models could crash or return stale results in several cases:

• whitespace-only chunks could reach the embedding path;
• encoder models could receive chunks larger than n_ubatch;
• llama’s process-global logger could retain stale per-engine state;
• changing token options after model initialization could leave the local engine sized for old chunk limits;
• cached embeddings could be reused after a local context rebuild even though tokenization/truncation may change.