Port merge_embeddings benchmark to tritonbench

[q10]

Summary:
Add Kineto trace export support to the merge_embeddings benchmark for A/B comparison with the tritonbench port.

This diff extends merge_embeddings_benchmark.py with two new CLI options that enable exporting Kineto Chrome traces. This is needed for kernel-level A/B comparisons between the existing FBGEMM benchmark and the new tritonbench port of the same operator.

1. --export-trace flag: When enabled, switches the profiler from the existing ProfilerActivity.CUDA + table-print path to a on_trace_ready callback that exports a Chrome trace file.
2. --trace-url template: Configurable output path with {ospid} placeholder for the PID, defaulting to merge_embeddings_fwd_trace_{ospid}.json.
3. _kineto_trace_handler() inner function: Formats the trace URL with the current PID and calls p.export_chrome_trace(url).
4. Plumbing: Both new options are threaded from the cli() Click entrypoint through to benchmark_merge_pooled_embeddings().

The existing (non-trace) profiling behavior is preserved as the default path when --export-trace is not passed.

Detailed Changes

merge_embeddings_benchmark.py

• Added import os and from contextlib import nullcontext (note: nullcontext is unused and flagged by lint)
• Added export_trace and trace_url parameters to benchmark_merge_pooled_embeddings()
• Refactored profiling block into two branches: trace-export path (on_trace_ready) vs. existing CUDA activity profiling path
• Added --export-trace and --trace-url Click options to cli()
• Plumbed new options through both the sweep loop and single-run call paths

Reviewed By: henrylhtsang

Differential Revision: D101068554

[meta-codesync]

@q10 has exported this pull request. If you are a Meta employee, you can view the originating Diff in D101068554.