#1752 HIP: unbuffered file-tier reads and a locked expert arena (#1691)

说明

Bisected the gfx906 prompt stalls and 0.1.41 verify timeouts to 805088d (full story in #1691).

**What happens without this, on 1x MI50 16 GB (gfx906), ROCm 10.0 and 10.1, 60 GB RAM, 43 GiB resident budget:**

- 805088d's 5%-of-read-bytes rule puts this box on cached file-tier reads; 0.1.40 required room for all read bytes, so it read unbuffered.
- Cached here means ~42 GB of storage I/O per prompt through a ~6.6 GB cache (3.7x amplification in the engine's own file-tier counters) - relentless reclaim.
- The 43 GiB arena is HMM-tracked but unpinned (VmPin 0 after hipHostRegister), so reclaim unmaps it and KFD suspends ALL GPU queues: up to 45 of 60 s idle, 7-26 s per prompt. Past 60 s the watchdog kills the engine; starved verify spin flags time out instead.
- Plain v0.1.40.4: first prompt dies, ~34 tok/s limping. Plain v0.1.40 (same box/config): 358-366 tok/s, 0.0 s suspended.

**This PR (HIP-only, 11 lines, one file):**

1. Keep the old room >= read_bytes rule on HIP (the #1194 5% rule stays everywhere else - your P100 measurement stands where reclaim has no queue side effects).
2. Lock the whole HIP-registered arena (mlock), as hardening: it stops the crashes alone but not the 2x slowdown, the cache churn itself costs that.

**Measured on the same box, v0.1.41 + this PR, full benchmark (idle, 3x ~4k prompts, 300-token story, 6 GiB RAM pressure, prompt after):**

| | before | after |
|---|---|---|
| prompt read | ~34 tok/s (then dies) | 316-366 tok/s |
| story generation | n/a | 27.3 tok/s |
| GPU suspended per prompt | 7-26 s | 0.0-0.1 s |
| watchdog stalls / verify timeouts | 1+ per session | 0 |

Reverting only 805088d on plain v0.1.40.4 gives the same numbers, so (1) is the fix and (2) is belt and braces. Happy to split them into two PRs if preferred.