[WebGPU EP] Add unit test for SessionBufferPool

## Background

PR #28761 introduces `SessionBufferPool` and the `sessionBufferPoolGenerations` provider option for the WebGPU EP, letting a session retain retired per-graph `BufferManager` buffer caches and seed them into newly created generators. End-to-end behavior is validated via `verify_multi_gen.py` and `verify_max_length_change.py` (Phi-4 on graph capture: `storage hits=171 misses=0, uniform hits=296 misses=0` after the first generator donates), but there is no dedicated unit test for the pool itself.

This issue tracks adding one as a follow-up. From [the PR review thread](https://github.com/microsoft/onnxruntime/pull/28761#issuecomment-4663803276):

> A unit test with a mock `IBufferCacheManager` exercising `Donate → SeedInto` roundtrip, `Donate` eviction at capacity, and `Clear` would lock in the contract and make the pool tractable to modify safely later.

## Proposed coverage

A new `session_buffer_pool_test.cc` (alongside the other WebGPU EP unit tests) with a fake `IBufferCacheManager` / `BufferManager` shim. At minimum:

1. **Donate → SeedInto roundtrip** — donate a slot with N storage + M uniform entries, immediately seed into a fresh manager, assert all entries land in the receiving cache with sizes and handles preserved.
2. **FIFO eviction at capacity** — with `max_generations = k`, donate `k + d` slots and assert that the `d` oldest slots' buffers were released (via `wgpuBufferRelease`) while the newest `k` remain pooled. The mock should track release counts to prove no leak.
3. **LIFO consume** — confirm `SeedInto` drains the most-recently-donated slot first; relevant for hot-reuse / shape stability.
4. **`max_generations == 0`** — `Donate` is a no-op (caches still extracted but immediately discarded with proper release on the retiring manager side).
5. **Empty-slot early-exit** — donating a manager whose caches extract to empty does not push a slot.
6. **`Clear` releases everything** — after `Clear`, the mock buffer release count equals the number of donated handles still pending; pool `Size()` returns 0.
7. **Destructor releases** — same property as `Clear` but via `~SessionBufferPool()`.

## Non-goals

- GPU-side fence behavior (covered by the documented invariants on `Donate` plus existing end-to-end tests).
- Threading (`SessionBufferPool` is documented as relying on `InferenceSession::session_mutex_` for serialization; no need to test concurrency).

## Notes

- Should land after the companion GenAI-side `State::~State() → SessionReleaseCapturedGraph` change so the feature is actually exercised in production. There is no urgency on this issue until the pool is in the steady-state hot path.
- The mock `IBufferCacheManager` only needs to implement `ExtractCachedBuffers` and `AbsorbCachedBuffers` — the rest of the cache interface is irrelevant for the pool's contract.