Flatten Megatron configs, smoke, and sweep cell keys and update docs

[sukesh-amd]

Summary

Document unified suites and Primus checkpoint save

• Move suite and config detail from the long test/config READMEs into published RST (docs/how-to/test-suites/training/megatron.rst and docs/reference/configuration-files/training/megatron.rst).
• Pass Primus --save on checkpoint writes so resume can load a real checkpoint.

Drive test_smoke from a top-level smoke block

• test_smoke reads smoke.enabled, iters, micro_batch_size, global_batch_size, and precision from the variant JSON instead of hardcoded knobs.
• enabled=false skips smoke. All packaged Megatron variant configs include this block.

Flatten Megatron configs onto paths, train_params, and container.env

• Drop nested config / model_params / schema_version / framework.
• Use gpu_name, paths, train_params, and container.env (NNODES, MASTER_ADDR, NCCL_*, sockets as docker run -e).
• Flatten those sections into the job dict at load time. Wrapper scripts no longer re-export NCCL/socket/MASTER_ADDR/NNODES.
• Threshold specs with "optional": true SKIPPED in test_metric when the metric is missing (training.mem_usage; distributed also training.scaling_efficiency_pct).
• Checkpoint I/O parser matches Primus loading distributed checkpoint from as well as Megatron-LM loading checkpoint from.
• Keep how-to and schema RST in sync with the flattened layout.

Require sweep keys to match MBS/GBS/PRECISION

• Rename every packaged sweep.combinations key and matching sweep.runs entry to MBS=<micro_batch_size>,GBS=<global_batch_size>,PRECISION=<precision> (same string as the threshold cell and the pytest parametrize ID).
• Config load fails if a combination key does not match those three fields, so pytest IDs and threshold lookup cannot drift.
• Log directories still sanitize = and , to _ (for example MBS_4_GBS_128_PRECISION_FP8).
• Update how-to and schema RST examples to the new keys.

Test plan

• make fmt-check lint ut
• Load a filled *_single.json and *_distributed.json via load_training_variant (all <changeme> replaced)
• Single-node: smoke from smoke.*; one sweep cell; optional mem_usage SKIPPED if the parser does not emit it
• Distributed Primus: checkpoint save/resume with --save; load log contains loading distributed checkpoint from
• Confirm combo keys are MBS=…,GBS=…,PRECISION=…; a leftover slug like llama3_1_8b-mi300x-… fails load

[solaiys]

Blocking: Primus save is a checkpoint directory, not a boolean flag. Bare --save will be parsed as the path --save_interval and the actual interval token is left dangling, so checkpoint write/resume still cannot work.

[solaiys]

this also should be <changeme> with example.
Either give the example in a comment line key:value.
OR
add the <changeme> tag at the end of the actual value what we are using

ex:
"NCCL_IB_HCA": "rdma0,rdma1,rdma2,rdma3,rdma4,rdma5,rdma6,rdma7 <changeme>",
"NCCL_SOCKET_IFNAME": "eno0 <changeme>",
"GLOO_SOCKET_IFNAME": "eno0 <changeme>",
"NCCL_IB_GID_INDEX": "3 <changeme>",

[sukesh-amd]

ok

[solaiys]

changeme tag should accompany with the example values.
submit it in the next PR.