Consolidate MFU tracking into perf_logger, address PR review feedback

Reworks MFU tracking per reviewer feedback on #1548:

- Delete per-recipe flops.py, test_flops.py, and the CLI entirely
- Inline ~30-line FLOPs helper into each recipe's existing perf_logger.py
- MFU metrics (train/tflops_per_gpu, train/mfu_pct) flow through the existing
  torchmetrics -> WANDB path, respecting logging_frequency
- Drop comm-overhead estimation; will be a separate future PR

The new formula is per_token_flops(seq_len) * num_unpadded_tokens_on_rank.
The unpadded-tokens counter (already used by tokens_per_second_per_gpu) is
per-rank after DP/CP sharding and accumulated across grad-acc micro-batches,
so the formula works uniformly across DDP/FSDP2/FSDP2+CP/DDP+CP/mFSDP and
across BSHD and THD (sequence packing) with no per-strategy factors.

Net: -3000 / +300 lines. Training scripts lose all MFU scaffolding; the
only change per script is one extra kwarg on the PerfLogger constructor.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Signed-off-by: Gagan Kaushik <gkaushik@nvidia.com>

Author: gagank1

bionemo-recipes/recipes/codonfm_native_te/README.md

@@ -185,14 +185,16 @@ Enable per-step Model FLOPs Utilization (MFU) logging during training by adding
 torchrun --nproc_per_node=1 train_fsdp2.py --config-name encodon_1b log_mfu=true
 ```
 
-This logs MFU (%), TFLOPS/GPU, and step time at each optimizer step. The module auto-detects model architecture from the model config.
+This adds two metrics at each logging interval, emitted alongside existing metrics via WANDB and
+stdout:
 
-The `flops.py` CLI provides standalone utilities:
+- `train/tflops_per_gpu` — achieved BF16 TFLOPS per GPU
+- `train/mfu_pct` — MFU as a percentage of the GPU's peak dense BF16 TFLOPS
 
-```bash
-python flops.py gpu-info                                       # Show GPU and peak TFLOPS
-torchrun --nproc_per_node=2 flops.py bandwidth                # Measure P2P GPU bandwidth
-```
+The FLOPs formula auto-detects model architecture from the model config (MHA, standard FFN,
+vocabulary size) and scales with the actual unpadded token count on each rank. This means it
+naturally handles gradient accumulation, data parallelism, BSHD, and THD (sequence packing)
+without per-strategy code paths. The implementation lives in `perf_logger.py`.
 
 ## Developer Guide