feature: add routed MoE compiler substrate

[Happymic]

Summary

Adds the compiler-side substrate for routed MoE execution and GPT-OSS MoE validation.

Changes

• Add routed MoE lowering helpers and GPT-OSS MoE program emission.
• Add compiler/ISA support used by dynamic addressing, vector min/max, top-k, tile-row helpers, and routed expert execution.
• Add GPT-OSS MoE reference utilities and correctness assertions.
• Update ISA documentation for the new/extended operations.

Scope notes

• Large experiment artifacts are not included.
• aten/gpt_oss_real_layer0_probe.py is intentionally not included; it is a local probe script, not part of the reusable compiler substrate.
• Simulator and RTL support are submitted as separate PRs.

Validation

Ran locally:

python -m pytest \
  aten/tests/test_plena_compiler.py \
  assembler/tests/test_vector_rmask_handling.py \
  aten/tests/test_gpt_oss_moe_assertions.py \
  aten/tests/test_gpt_oss_moe_reference.py

[booth-algo]

LGTM