Validate rampup batch size settings in the Pydantic config

[Atishyy27]

Closes #4690.

The legacy config validates the rampup batch size settings in pyconfig_deprecated.validate_rampup_batch_size (src/maxtext/configs/pyconfig_deprecated.py:203). The Pydantic config has no equivalent, so an unusable combination is accepted and then silently degrades.

The derivation in MaxTextConfig is guarded by if self.global_batch_size_to_load_increment > 0: and if num_increments > 0:, so a non-positive increment, non-positive global_rampup_samples, or a start at or above per_device_batch_size falls through and leaves rampup_end_step = 0. Rampup is off and nothing says so. When the change is not a multiple of the increment, num_increments = diff // increment truncates and ramp-up finishes below the configured per_device_batch_size.

Replicating the derivation arithmetic for 8 devices, expansion_factor_real_data=1, gradient_accumulation_steps=1:

settings	rampup_end_step	result
per_device_batch_size=8, start=4, increment=2, samples=500	14	correct
increment=0	0	rampup silently off
global_rampup_samples=0	0	rampup silently off
start=8, per_device_batch_size=4	0	rampup silently off
per_device_batch_size=9, start=4, increment=2	14	ramps to global batch size 64, 72 requested

This adds a model_validator(mode="after") on DatasetGeneral with the same five checks as the legacy path, skipped entirely when enable_rampup_batch_size is False.

Tests in tests/unit/configs_value_test.py cover each rejection, a positive control asserting a well-formed config still produces rampup_end_step > 0, and a case confirming the checks do not fire when rampup is disabled.

[gemini-code-assist]

Caution

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