[REFACTOR][GEMM] Merge standalone gemv into gemm and migrate gemv tests

[RMLYC]

Problem

gemv and gemm are implemented and tested as separate operator stacks, even though GEMV is a shape-specialized GEMM case.

Affected files/functions:

• tileops/ops/gemv.py (GemvOp) duplicates operator dispatch and lifecycle already present in tileops/ops/gemm.py (GemmOp).
• tileops/ops/__init__.py exports both GemmOp and GemvOp, keeping two public APIs for one conceptual op family.
• tests/ops/test_gemv.py and tests/ops/test_gemm.py are split, so GEMV coverage is disconnected from GEMM regression checks.

Problematic pattern today:

# tileops/ops/__init__.py
from .gemm import GemmOp
from .gemv import GemvOp

__all__ = [
    ...,
    "GemmOp",
    "GemvOp",
    ...,
]

And separate op entrypoints:

# tileops/ops/gemv.py
class GemvOp(Op):
    ...
    self.kernel = self.kernel_map["gemv_kernel"](n, k, self.dtype, tune=tune)

# tileops/ops/gemm.py
class GemmOp(Op):
    ...
    self.kernel = self.kernel_map["gemm_kernel"](
        m, n, k, self.dtype, tune=tune, trans_a=trans_a, trans_b=trans_b
    )

Proposed Fix

Unify GEMV into GEMM as a shape-specialized path and remove standalone GEMV ops/test entry.

Implementation plan:

1. Remove standalone GEMV op exports and entrypoints (GemvOp, and related imports).
2. Route GEMV workloads through GemmOp by expressing GEMV as GEMM shape mapping (e.g., M=1 / equivalent shape transform in the operator wrapper).
3. Migrate all tests/ops/test_gemv.py coverage into tests/ops/test_gemm.py (or GEMM test matrix), preserving dtype/tuning/shape scenarios from current GEMV tests.

Target end-state sketch:

# only GEMM-facing public API
from .gemm import GemmOp

# GEMV cases executed through GEMM-compatible shapes
op = GemmOp(m=1, n=n, k=k, ...)
out = op(a_2d, b)

Next Steps

A follow-up PR will implement the refactor with:

• API cleanup (remove standalone GEMV surface).
• Kernel dispatch unification under GEMM.
• Test migration table (old GEMV case -> new GEMM case) included in PR description.
• Correctness verification for migrated cases and a brief performance sanity check on representative GEMV-shaped inputs.