feat(moderncolbert): batched POST /pooling support in IO processor

[ddickmann]

Summary

Adds opt-in batched-input support to the moderncolbert_io IOProcessor. A single POST /pooling call can now carry N texts and the processor decomposes them into N prompts inside one factory_pre_process step, letting vLLM's continuous batcher co-schedule them in one engine pass.

Why

Surfaced while wiring moderncolbert_io into a high-fan-out groundedness scorer where one request typically encodes ~30 chunks with the same model. With the existing one-text-per-call contract, the choice was either:

• N concurrent HTTP calls per request (CPU-bound on syscalls + serialization, blocks the engine batcher from co-scheduling them), or
• A serial latency floor.

Single-call batching collapses both problems and gives the engine batcher the freedom to fuse prompts within a request as well as across them.

Wire format

The single-text shape is unchanged and byte-identical to the previous release — same request body, same single-string base64 response — so existing callers are not affected.

The new batched shape is opt-in:

{
  "data": {
    "text": ["q1", "q2", "..."],
    "is_query": [true, true, true]
  },
  "model": "lightonai/LateOn",
  "task": "plugin"
}

• is_query may also be a single bool that broadcasts across all texts.
• Mixed-type lists raise an explicit error so callers don't silently mis-pack a batch.
• Empty batches raise.
• The response for batched requests is list[str] of base64-encoded multi-vector embeddings, in input order.

Implementation notes

• ModernColBERTInput now carries a list of texts plus a per-text is_query list and a batched flag recording the request-level shape.
• factory_parse detects single vs. batched input from the type of data['text'] and validates is_query length consistency.
• factory_pre_process returns Sequence[TokensPrompt] for batches and a single TokensPrompt for the legacy path (preserving the upstream prompt-shape contract).
• factory_post_process returns str for legacy single-text requests and list[str] when the request was batched (or when multiple outputs come back), keyed off the stashed batched flag rather than relying on len(model_output) alone — that way batches of size 1 still round-trip as list[str].

The [Q] / [D] prefix logic, max-length budgets (256 query / 8192 doc), pooling task name, and tokenizer settings are all unchanged.

Backward compatibility

Caller shape	Before	After
{"text": "abc", "is_query": true}	str (b64)	str (b64) ✅
{"text": "abc", "is_query": [true]}	error	str (b64)
{"text": ["abc"], "is_query": true}	error	["..."]
{"text": ["a","b"], "is_query": [true,true]}	error	["...","..."]

Single-text path is fully preserved; lists are a strict superset.

Test plan

• CI passes
• Single-text request still returns a str response (regression check vs. existing client code)
• Batched request with N texts returns list[str] of length N in input order
• Mixed-type lists in text raise ValueError
• Length-mismatched is_query list raises ValueError
• Empty text batch raises ValueError
• Bool is_query broadcasts across a list of texts