Add --profile-json to run/run-c/run-cu: machine-readable run output

[SiliconState]

Problem

Every tool that measures Bend today (CI benchmarks, profilers, editors) re-parses the human output:

Result: 16744448
- ITRS: 5767113
- TIME: 0.06s
- MIPS: 98.01

That format is undocumented and implicitly frozen: any change to the Result: prefix, the - KEY: value stat lines, or the 0.06s unit breaks every downstream parser silently. Each consumer writes (and re-debugs) the same fragile scraper. The scraping contract already exists — it is just informal; this PR makes it explicit and cheap to keep, instead of frozen-by-fear.

Design

A --profile-json flag on run, run-c, and run-cu that replaces the Result:/stats lines with one compact JSON object on stdout:

$ bend run-c prog.bend --profile-json
{"backend": "run-c", "result": "16744448", "stats": {"itrs": 5767113, "time_s": 0.06, "mips": 98.01}}

• backend: which runner produced the measurement, in Bend's own naming (run-rs, run-c, run-cu; HVM internally calls run-rs just run).
• result: the readback term, exactly the text that followed Result: .
• stats: the HVM - KEY: value lines parsed into typed fields. Keys are lowercased; numeric values become JSON numbers; TIME's trailing s unit is stripped and the field named time_s. Unknown keys pass through, so new runtime stats appear automatically without a Bend release.
• Program output (from IO) still streams to stdout during the run, so tools take the last line of stdout. Diagnostics stay on stderr.
• No new dependencies: emission is a small formatter with a JSON string escaper. Values are emitted as JSON numbers only when they match the strict JSON number grammar — HVM prints - MIPS: inf when the reported duration is 0.00, and Rust's float parser accepts inf/NaN/+5/.5, none of which are valid JSON, so those are quoted instead.

Patch (src/main.rs only)

1. CliRunOpts gains --profile-json.
2. The run arm prints profile_json_line(backend, &term.to_string(), &stats) instead of the Result:/-s lines when the flag is set (-s/-p are subsumed: stats are always included, and the result is the plain term text).
3. profile_json_line + is_json_number + json_escape helpers with unit tests (typed stats parsing, strict-number quoting, escaping, non-stat lines skipped).

Backward compatible: without the flag, output is byte-for-byte unchanged.

Testing

• Unit tests for stats parsing (itrs/time_s/mips/unknown-key passthrough), JSON number strictness, string escaping, non-stat-line skipping.
• End-to-end: bend run-c parallel_sum.bend --profile-json emits one valid JSON line (validated with python3 -m json.tool) whose itrs matches the -s run; run-rs works identically; full test suite passes.

Composition

With a runtime thread count (#763 + HigherOrderCO/HVM2#441), a complete machine-readable scaling experiment is one shell line:

for n in 1 2 4 8; do bend run-c prog.bend --threads $n --profile-json; done