[test] Add me2e pytest framework for middle end

[brnorris03]

What?

New pytest-based ME2E (middle-end to end-to-end) test framework in test/me2e/
for testing TTL MLIR generation, compilation, and hardware execution.

Why?

• Unified framework for testing TTL MLIR generation, compilation, and hardware
  execution
• Declarative parametrized tests enable testing all ops with all configs via
  single test function
• Auto-generates test classes from TTLElementwiseOps.def keeping tests in sync
  with dialect
• ULP-based comparison utilities for hardware-accurate golden validation
• Supports custom fused op tests via MLIR templates

How?

Two test approaches (both auto-generated from TTLElementwiseOps.def):

1. Declarative parametrized tests (primary): Single test_compute_ops.py
   function parametrized over all operations and configurations. COMPUTE_OPS
   in op_specs.py is auto-generated from the .def file. All pipeline stages
   executed in one function via runner.py. Fast, clean, no artifacts.

2. Class-based tests: Test classes auto-generated from .def file. Classes
   declare what to test (op name, arity, input range, tolerance) via class
   attributes; base classes handle how (MLIR generation, compilation,
   execution, validation). Saves intermediate (MLIR, CPP, results) artifacts
   for debugging, supports custom (fused) ops via MLIR templates.

Architecture:

• E2ETestBase: Ordered 5-stage pipeline (build → compile → translate → execute
  → validate)
• OpTestBase: Adds OP_STR, ARITY, INPUT_RANGE; auto-generates from
  .def file
• FusedOpTestBase: Custom MLIR via get_mlir_template() override
  (examples in test/me2e/ops/test_fused.py)
• Host configuration and execution reuses the shared kernel_runner
  module (python/ttl/kernel_runner.py)
• Artifacts saved to build/test/me2e/<TestName>/ for debugging (class-based
  tests) or temporary directories (declarative tests).

flowchart TD
    subgraph Inputs
        DEF[.def file]
        MLIR[Custom MLIR]
    end

    DEF --> GEN[Generate: TestAdd, TestExp...]
    DEF --> DECL[Declarative: test_compute_ops]
    MLIR --> FUSED[FusedOpTestBase]

    GEN --> P1
    FUSED --> P1
    DECL --> P2

    subgraph P1[Class-based Pipeline]
        S1[build] --> S2[compile]
        S2 --> S3[translate]
        S3 --> S4[execute]
        S4 --> S5[validate]
    end

    subgraph P2[Declarative Pipeline]
        D1[build] --> D2[compile]
        D2 --> D3[translate]
        D3 --> D4[execute + validate]
    end

    S1 --> O1[module.mlir]
    S2 --> O2[compiled.mlir]
    S3 --> O3[*.cpp]
    S4 --> O4[result.pt]

    D4 --> O5[in-memory validation]

Loading

Current status:

• All 5 pipeline stages (build, compile, translate, execute, validate): Working
  for all ops
• Tests the 13 elementwise operations defined in TableGen (add, sub, mul, max, exp, log, sqrt, rsqrt,
  tanh, abs, neg, relu, sigmoid)
• Two grid configurations tested: 1x1 (single tile) and 2x2 (4 tiles) via parametrization
• All single op tests are parameterized by dtype (bf16 and f32) but some f32 tests' results have error exceeding the threshold and are marked with xfal): [BUG] Wrong results for add/sub/log elementwise ops with f32 #254

See
https://github.com/tenstorrent/tt-lang/blob/bnorris/me2e-tests/test/me2e/README.md
for more details.

How to Test?

Part of the check-ttlang-all target. Also separately testable with:

source build/env/activate
pytest test/me2e/ -v              # All me2e tests
ninja -C build check-ttlang-me2e  # CMake target

CI:
https://github.com/tenstorrent/tt-lang/actions/runs/21126783459/job/60749650887

Checklist:

• Self-reviewed (style, logic)
• Added tests (227 new tests)
• PR is focused (ME2E test framework only)