fix(#178): complete chunked prefill — KV budget cap, FlashInfer workspace, ragged attention

[drunkcoding]

Completes chunked prefill (#178). Stacks on #189 (fix/serving-streaming-and-flashinfer-dtype); merge #189 first.

Fixes

1. KV block-budget reconciliation — the logical KV allocator and the physical paged store were budgeted independently, so the logical count could exceed the store; _bind_layered_paged_kv_store then failed its capacity gate and swallowed the error, leaving block_store unbound and crashing _schedule_chunked_prefill ("layered paged KV store is not bound"). Now caps logical→physical before binding (PagedKVCache.resize_num_blocks), guards per-sequence decode capacity, and logs the real reason instead of a bare except.
2. FlashInfer workspace — chunked append-attention split-KV scratch overflowed the 128 MiB default on chunk 2; raised the shared prefill+decode workspace to 256 MiB.
3. Ragged query packing — concurrent requests with heterogeneous query lengths padded every sequence to q_len, feeding padding tokens to the paged backend; now packs valid tokens per query_lengths and scatters results back.

Verification (Qwen3-30B-A3B, sm120, FlashInfer 0.6.18)

• Chunked prefill runs end-to-end: short + 1000-tok (4 chunks) + 4096-tok (16 chunks) + decode.
• Whole-prefill non-regressing (store binding now also covers it).
• A/B (benchmarks/serving/chunked_prefill_latency.py) passed with output_parity=True (178_ab.json).

Known pre-existing limitation (out of scope)

Whole-prefill hangs on prompts > max_tokens_per_step (2048) — the exact case chunked prefill solves; tracked separately.