GLiNER-family adapters: hardening follow-ups from #380 and #381 #382

Hardening follow-ups for the GLiNER-family adapters, from the review of #380 and #381. None of them block those PRs.

• GLiNER2 caller errors should raise InvalidInputError. (Completed in fix(server): report GLiNER2 caller errors as invalid input #526.) Bad metadata.entities on relations, invalid labels or schema, and multi_label misuse currently raise ValueError, which surfaces as a server error. Raise InvalidInputError instead, as GLiREL now does, so these return 400 INVALID_INPUT.
• Count GLiNER2's real prompt. (Completed in fix(gliner2): count compiled inference prompts #794.) Measure the prompt gliner2 actually builds, as the GLiNER path does, instead of estimating it. The estimate over-counts plain field specs, so some valid schemas are rejected, and it can under-count JSON-heavy ones by about 2%.
• Validate the prompt before windowing the document in GLiNER2. (Completed in fix(server): validate GLiNER2 prompts before document work #793.) Today the document is windowed and metered first. That work is already bounded, so this is an efficiency change.
• Apply the 128-character label check to the GLiNER bi-encoder. (Completed in fix(server): reject overlong GLiNER bi-encoder labels #651.) This matches the other GLiNER-family adapters.
• GLiNER2 metering covers the window. (Completed in fix(server): meter compiled GLiNER2 document windows #797.) GLiNER2 meters document tokens up to 512, while the window it reads can hold up to 2048 subwords. Consider metering the window's subwords, as GLiNER does.
• Docstring. (Completed in fix(server): validate GLiNER2 prompts before document work #793.) Update the 50-field schema figure in _prompt_limit.py to about 1,260 tokens.