[Feature] Add offline training operator profiling (#1310)
Summary

decouple operator report generation from the training profiling
callback so training only exports rank-scoped Chrome timelines
add a streaming offline extractor with External ID attribution,
explicit mapping status, communication accounting, training operator
hierarchy, multi-rank batch extraction, and cross-rank aggregation
align per-rank outputs with the inference profiler:
operator_list.csv, kernel_time.csv, kernel_shape_dtype.csv, and
summary.json, using a rank-N_ prefix
retain unattributed kernels and communication kernels, use all GPU
kernel duration as the timing denominator, and validate
event/time/CSV/ID conservation
add unit coverage and a training operator profiling Quick Start

Output compatibility
The old synchronous report names are replaced by:

rank-N_operator_list.csv
rank-N_kernel_time.csv
rank-N_kernel_shape_dtype.csv
rank-N_summary.json

Communication is included in the common reports instead of a separate
communication CSV. Training-specific hierarchy fields are appended after
the inference-compatible columns.
Testing

python -m pytest -q tests/unit_tests/train/test_train_config.py tests/unit_tests/train/utils/test_optim_setup.py tests/unit_tests/train/megatron/test_profiler_reports.py (86 passed)
python -m compileall -q flagscale/train/megatron/training/profiler_reports.py flagscale/train/megatron/training/training.py
git diff --check
validated both fusion-on and fusion-off real training timelines; all
conservation checks passed