feat(nnx): fix checkpoint restore ShapeDtypeStruct crash under native LoRA

[RexBearIU]

Description

This PR fixes a JAX ShapeDtypeStruct crash during checkpoint restoration under Flax NNX native LoRA (and QLoRA).

IMPORTANT NOTE: This pull request is blocked by PR #4346 (feat(nnx): support native Flax NNX PEFT/LoRA training loop). It targets main directly to keep its commit history clean, but should only be merged after PR #4346 has been successfully integrated into main.

Key Changes

• Treats Fallback SDS as Absent: Fallback ShapeDtypeStructs from Orbax partial restores are treated as absent/None to prevent them from overwriting concrete initialized weights.
• Materializes Template State: Properly materializes and shards the concrete template state using init_state_fn before overwriting restored weights.
• Refactored Integration Tests: Transitions the PEFT integration tests to use sequential isolated Python subprocesses. This completely avoids JAX exclusive-locking conflicts in multi-device/TPU VM host environments, allowing the test suite to execute successfully on actual accelerator hardware.
• Dynamic Multi-Device TPU Support: Dynamically detects the TPU VM device count and overrides ici_fsdp_parallelism=1 and ici_data_parallelism=num_devices for downscaled test models, eliminating any compilation FSDP partitioning divisibility constraints.

FIXES: #4400

Tests

The complete sequential native LoRA integration test suite covering Gemma4 and Qwen3 SFT, and LLaMA2 pre-training save/restore loops was successfully run and verified on both CPU/CI and real TPU VM host environments:

PYTHONPATH=src pytest tests/integration/peft_integration_test.py

All 6 integration tests successfully pass:

• test_native_sft_gemma4_lora (passed)
• test_native_sft_gemma4_qlora (passed)
• test_native_sft_qwen3_lora (passed)
• test_native_sft_qwen3_qlora (passed)
• test_pretrain_lora_checkpoint_save_and_restore (passed)
• test_pretrain_qlora_checkpoint_save_and_restore (passed)

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have run end-to-end tests tests and provided workload links above if applicable.
• I have made or will make corresponding changes to the doc if needed, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.

[codecov]

Codecov Report

❌ Patch coverage is 63.41463% with 30 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...rc/maxtext/utils/generate_param_only_checkpoint.py	73.21%	8 Missing and 7 partials ⚠️
src/maxtext/utils/standalone_checkpointer.py	0.00%	6 Missing ⚠️
src/maxtext/common/checkpointing.py	50.00%	2 Missing and 3 partials ⚠️
src/maxtext/utils/maxtext_utils.py	55.55%	3 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!