fix(megatron): restore "compute per GPU" TFLOP label and make log parsers robust

[amd-fuyuajin]

Summary

Restores the compute per GPU (TFLOP/s/GPU) training-log label (from #842) that
was inadvertently reverted by the Flux MLPerf logging refactor (#820), and updates
the three Megatron log parsers so both the current and legacy log formats parse
correctly.

Background

• improve the logging format for the megatron backend #842 relabeled the per-GPU FLOP rate from throughput per GPU (TFLOP/s/GPU): X/Y
  to compute per GPU (TFLOP/s/GPU): X (avg Y). "Throughput" is ambiguous next to
  the tokens/s line; compute describes a FLOP rate unambiguously.
• feat(flux): MLPerf logging/warmup/lr-schedule patches #820 extracted a shared _inject_tflops() helper for the diffusion path and, in
  the process, reverted that label back to throughput per GPU. The harmonic-mean
  token change from improve the logging format for the megatron backend #842 was left intact.
• The newest docs (examples/odc/README.md, added after feat(flux): MLPerf logging/warmup/lr-schedule patches #820) already document the
  log as compute per GPU (TFLOP/s/GPU), so the code had drifted from the docs.

What changed

primus/backends/megatron/patches/training_log/print_rank_last_patches.py

• Restore the compute per GPU (TFLOP/s/GPU): X (avg Y) label in _inject_tflops().
• Refresh the module/render docstrings to match the actual output
  (compute per GPU ... and tokens/s/GPU inst/harmonic mean: ...).

Log parsers — tools/daily/daily_report.py, tools/auto_benchmark/metrics.py,
benchmark/megatron/model/benchmark_report.py

• Accept both the current labels (compute per GPU, tokens/s/GPU inst/harmonic mean)
  and the legacy ones (throughput per GPU, tokens per GPU (tokens/s/GPU)), using
  non-capturing alternations so positional group indices are unchanged.
• benchmark_report: the memory segment is now appended last (as
  hip mem usage/free/total/usage_ratio: X GB/...) rather than sitting between
  elapsed and throughput as mem usages: X; reorder the pattern/group indices and
  accept both memory labels.
• Tolerate the optional space before (avg (real logs can render X(avg Y)), and
  fall back to the instantaneous value when a running average is absent (raw
  warmup lines).

Why the parsers were already broken

The token label from #842 (tokens/s/GPU inst/harmonic mean) never matched the
parsers' tokens per GPU (tokens/s/GPU) regex, and benchmark_report expected a
mem usages: segment that current logs no longer emit. This PR realigns all three
with the current log format while staying backward-compatible.

Test plan

• pytest tests/unit_tests/backends/megatron/test_training_log_patches.py — 7 passed
• End-to-end: drove the real ThroughputAverageExtension/ElapsedAverageExtension
  across multiple iterations, rendered the line, and confirmed all three parsers
  extract correct inst/avg TFLOP, tokens, and memory values.
• Validated all three parsers against real pre-train logs
  (llama2_7B-BF16-pretrain, MI300X): current + legacy + raw-warmup lines all parse.
• No new lint errors.