V1 engine breaks block size for gpt-oss-120b

Possible regression / changed requirement in V1 XPU path:

Previously, `--block-size 32` worked with this model/config. On the current Intel Scaler / V1 XPU path, the server starts successfully, loads the model, completes init, then fails on first request with:

RuntimeError: XeTLA ChunkPrefill FP8KV: only support block_size >= 64

This is surprising because the config is accepted at startup and only fails at execution time after a long model load/warmup.

Observed path:
- vLLM V1 engine
- XPU backend
- GPT-OSS-120B
- MXFP4
- chunked prefill enabled
- Flash Attention backend
- `--block-size 32`

Workaround:
- `--block-size 64` works
- Next test: `--block-size 32 --disable-chunked-prefill`

Request:
Please document this requirement, reject invalid `--block-size` values at startup when this backend path is selected, or allow the old behavior if there is a supported non-FP8KV/chunked-prefill path.

This is especially painful because the failure happens after the model is already loaded and the API server has started.

A startup validation error would save a lot of debugging time. And reboots because it wedges the GPUs