feat: support epoch-based checkpoint intervals

[tongchen126]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

Add the --save_epochs training argument for epoch-based checkpoint intervals.

When set to a positive integer N, save_epochs automatically selects epoch-based checkpointing and saves at the
end of epochs N, 2N, and so on. The option is supported by standard Transformers-based Swift training, native
Megatron training, and Ray Megatron training. Epoch-based evaluation follows the same interval so that
load_best_model_at_end and best-checkpoint tracking remain consistent.

The implementation validates that save_epochs >= 1, converts epoch intervals to step intervals for Megatron and Ray
Megatron, and rejects streaming datasets in native Megatron and Ray Megatron when save_epochs is requested because
their epoch length is unknown.
Existing save_steps and save_strategy behavior remains unchanged when save_epochs is not set. The training UI and
English/Chinese command-line documentation were updated accordingly.

Experiment results

• Python compilation and git diff --check passed for all modified files.
• Isolated behavior checks passed for HF save/evaluation cadence, Ray step conversion, streaming rejection, and
  compatibility with existing Ray save_strategy='epoch' configurations.
• A focused regression test was added. It could not complete in the available environment because the installed
  dependency stack hung during modelscope import; no test failure was reported.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.