Keep the elasticity batch overrides out of the caller's config dict

[alanhuangyoo]

DeepSpeedConfig keeps the dict it is handed by reference:

if isinstance(config, dict):
    self._param_dict = config

#8289 established that parsing must not write back into it — the caller owns that dict and may reuse it after initialization.

The elasticity branch still does, two lines above the comment that says otherwise:

        self._param_dict[TRAIN_BATCH_SIZE] = final_batch_size
        self._param_dict[TRAIN_MICRO_BATCH_SIZE_PER_GPU] = micro_batch_size
        self._param_dict[GRADIENT_ACCUMULATION_STEPS] = gradient_accu_steps

    # Pass a copy so that user json is unmodified, e.g. for logging
    self._initialize_params(copy.copy(self._param_dict))

A caller that enables elasticity gets three keys back that it never set. print_user_config() dumps self._param_dict, so it then reports them as though the user had written them.

The elasticity path also forbids those keys in the input unless ignore_non_elastic_batch_info is set:

One or more batch related parameters were found in your ds_config (...).
These parameters *will not be used* since elastic training is enabled ...

so a config that was rejected on the first pass would be accepted on a re-parse of the same dict, since the second time around the keys are there.

The fix

Collect the overrides and apply them to the copy. All three are top-level keys, so the existing shallow copy keeps them off the caller's dict.

Test

config_dict = {"elasticity": {"enabled": True, "max_train_batch_size": 4,
                              "micro_batch_sizes": [1, 2], "min_gpus": 1, "max_gpus": 4,
                              "min_time": 0, "version": 0.1,
                              "ignore_non_elastic_batch_info": True}}
keys_before = set(config_dict)
ds_config = DeepSpeedConfig(config_dict)

before

caller dict gained: ['gradient_accumulation_steps', 'train_batch_size', 'train_micro_batch_size_per_gpu']
parsed:             train_batch_size=4  micro=2  gas=2

after

caller dict gained: nothing
parsed:             train_batch_size=4  micro=2  gas=2

Same parsed values, so this only removes the write-back.

Added test_elasticity_leaves_caller_config_untouched next to #8289's test_max_grad_norm_leaves_caller_config_untouched. It fails on master with

AssertionError: assert {'elasticity', 'train_batch_size', 'train_micro_batch_size_per_gpu',
                        'gradient_accumulation_steps'} == {'elasticity'}

and passes with the fix.

tests/unit/runtime/test_ds_config_dict.py   27 passed, 5 skipped
tests/unit/elasticity/test_elastic.py       23 passed, 3 skipped
yapf --diff / flake8                        clean