[DRAFT] Allocate host staging per transfer instead of fixed slots

[yinlin09]

What

Allocates host staging per transfer instead of carving the pool into fixed
slots, behind TPU_RAIDEN_DYNAMIC_HOST_STAGING=1.

Today the staging pool is pre-partitioned into num_slots slots of
max_blocks blocks, and a transfer borrows a whole slot however few pages it
needs:

• slot size derives from the configured context length, so short requests idle
  most of their slot;
• the slot count is a hard cap on concurrent transfers, on both the consumer
  (StartRead) and the producer serving a pull (StartPushInternal);
• a producer that cannot get a slot marked the send done without
  transferring, leaving the consumer to burn its full timeout before failing.

How

Under the flag, both paths allocate exactly the pages they stage from the
existing host block manager (locked against LRU eviction for the transfer's
lifetime) and hand them back on completion, timeout, or error through the same
release helpers that returned fixed slots. The copy plans already address host
blocks explicitly, so nothing downstream changes and nothing changes on the
wire.

Exhaustion behaviour changes in both modes: the producer now waits for staging
to free (bounded by the transfer deadline) instead of fabricating completion,
and a request that can never fit is reported as a transfer failure with a
reason.

Off by default; the fixed-slot layout is untouched without the flag. The
reshard/pool-plan paths are unchanged.

Status — draft

Both wheel legs build green through ci/build_wheel.sh (jax and torch; the
torch wheel's three ABI extensions each carry the new code path). On-TPU
validation is in progress: a prefill/decode pair serving with the flag on, at
concurrency above what the fixed-slot layout could seat. Known follow-up:

• fragmentation under sustained mixed-size load reduces copy coalescing; needs
  a throughput comparison against the slot path at equal concurrency.