[Runtime][Python] Keep all results in benchmark_module (#24985)

`benchmark_module` parsed the console table by splitting whitespace
columns. It kept only the first user counter and failed on rows without
counters. Read a separate Google Benchmark JSON output file and expose
each complete row through `BenchmarkResult.report`, including metadata
and every user counter, without maintaining a list of Google Benchmark
fields.

This replaces `user_counters`. The name, timing, and iteration fields
retain their string representation. Percentage aggregates use `%`, and
an unmatched benchmark filter returns an empty list. Subprocess failures
and JSON rows marked `error_occurred` raise `BenchmarkToolError` with
their diagnostics.

The wrapper also accepts an explicit tool executable, benchmarks
eligible exports of a module file when no entry function is given,
supports repeated flags through list values, and serializes booleans as
`true`/`false`. These support the lifecycle tests in the following
changes.

Assisted by AI.

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