Refactor DeepSeek MoE/Engram scanned decoder tests to use 3 layers

[Shuwen-Fang]

Description

Context:
Large rematerialized DeepSeek V4 MoE/Engram scanned decoder tests (test_decoder_init_engram_2_8 and test_decoder_init_engram_4_9) experienced SPMD compilation buffer OOM crashes on 8-core TPU7x slices. This was due to memory pressure from unrolled chunks and workspace doubling on TPU7x (which lacks MegaCore, meaning 2 JAX devices share the same physical chip's HBM).

Solution:
Refactored the tests to use base_num_decoder_layers=3 instead of 5 and 4. This reduces the generated SPMD HLO graph size sufficiently to avoid OOM on TPU7x, while still validating the core scan-chunking compilation logic (Scan -> Engram -> Scan).
Also removed the temporary skip guard that was bypassing these tests on TPU7x.

Bug: b/527535999

Tests

Please describe how you tested this change, and include any instructions and/or
commands to reproduce.

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

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!