[iree-benchmark-module] Add dispatch timing rows (#24987)

Add `--dispatch_statistics` to report a per-function breakdown of
dispatch time beneath each `iree-benchmark-module` benchmark. Rows
include calls per model iteration and each function's share of total
dispatch time, across participating devices and benchmark repetitions.
They appear in console, JSON, and CSV output, including
`--benchmark_out` files and the Python wrapper's results.

After each measured repetition, Google Benchmark reruns the registered
benchmark outside the measured interval with a HAL statistics session on
every device. Profiling uses the same function ABI, batch accounting,
and asynchronous concurrency as measurement. These extra invocations can
advance model state. Output processing retains the measured pass's
outputs, including snapshots of buffers that subsequent invocations can
overwrite.

Conflicting device profiling, capture, and replay flags are rejected
before sessions start or replay output is opened. The tooling profiling
session API accepts explicit options, and benchmark-module device
profiling covers all participating devices.

Example with the included dispatch-statistics fixture:

```sh
iree-compile \
  --iree-hal-target-device=local \
  --iree-hal-local-target-device-backends=llvm-cpu \
  tools/test/iree-benchmark-module-dispatch-statistics.mlir \
  -o /tmp/dispatch-statistics.vmfb

iree-benchmark-module \
  --module=/tmp/dispatch-statistics.vmfb \
  --device=local-task --task_topology_group_count=2 \
  --function=main \
  --input=64x64xf32=1 --input=64x64xf32=1 --input=64x32xf32=1 \
  --benchmark_min_time=0.1s --time_unit=us --dispatch_statistics
```


Example output from a macOS arm64 run with two `local-task` workers
(timings vary):

```text
------------------------------------------------------------------------------------------------------
Benchmark                                            Time             CPU   Iterations UserCounters...
------------------------------------------------------------------------------------------------------
BM_main/process_time/real_time                    23.1 us         35.3 us         6181 items_per_second=43.226k/s
BM_main/main_dispatch_0_matmul_64x64x64_f32       8.17 us         12.1 us         6181 calls=2 percent=82.6871
BM_main/main_dispatch_2_matmul_64x32x64_f32       1.71 us         3.21 us         6181 calls=1 percent=17.3129
```

The 64×64 dispatch runs twice per model iteration, so its `Time`
includes both calls. `percent` is the share of total dispatch time; the
first row reports end-to-end model latency. On this CPU backend, the
dispatch rows' `CPU` column sums worker time across dispatch tiles.

Assisted by AI.

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>