fix: prevent OOM in embedding backfill with long texts

[BYK]

Summary

• Add truncation: true to the transformers.js pipeline call, capping individual texts at the model's max token length
• Reduce BACKFILL_CHUNK_SIZE from 32 to 8, limiting peak tensor size when long texts are batched
• Detect OOM errors (opaque ONNX numeric codes like 287180544) and return human-friendly messages with batch diagnostics
• Upgrade backfill error logging from log.info to log.error so failures are captured by Sentry

Root Cause

Nomic v1.5 pads ALL texts in a batch to the longest sequence length. A batch of 24 texts where one has 1383 tokens creates a [24, 1383] input tensor — the resulting intermediate activations cause a ~287 MB allocation failure in onnxruntime. The error was reported as an opaque numeric code "287180544" with no human-readable context.

Fix Details

Change	File	Effect
truncation: true	embedding-worker.ts:170	Caps each text at model max (8192 tokens), preventing any single oversized input
BACKFILL_CHUNK_SIZE = 8	embedding.ts:1097	Limits batch×sequence tensor size; worst case [8, 8192] vs previous [32, 8192]
isOomError() helper	embedding-worker.ts:166	Detects numeric-only error codes (≥6 digits) and OOM patterns, wraps in descriptive message
log.error in catch blocks	embedding.ts:1140,1197	Sends to Sentry via captureException (was log.info which skips Sentry)