[local-task] Clamp file read/write lengths to the platform transfer limit (#24816)

currently loading parameter files with "file" mode fails with:

```
IndexError: Error creating vm context with modules: external/+_repo_rules4+iree/runtime/src/iree/hal/drivers/local_task/task_queue.c:2714: OUT_OF_RANGE; short read: requested 13282859520 bytes, got 397957632; while invoking native function hal.fence.await; while calling variadic import;
```

The io_uring and IOCP backends narrowed the span length to their native
transfer count (sqe->len is 32 bits, ReadFile/WriteFile take a DWORD).
Spans over 4GB transferred the wrong amount, and a span that is an exact
multiple of 4GB wrapped to zero: the operation completed successfully
having moved no bytes, which the caller cannot distinguish from EOF.

Clamp instead. Short transfers are part of the file read/write contract,
so callers resubmit for the remainder. The ceiling is INT_MAX rounded
down to a page, matching the kernel's MAX_RW_COUNT.

Adds a CTS test covering the exact-4GB case, using a sparse reservation
so the oversized span costs address space rather than committed memory.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Signed-off-by: Thomas Ziereis <ziereis@roofline.ai>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>