Extend DotGeneralSimplify to handle broadcasted scalars

[binoygeorge97]

Fixes #1475.

Problem

DotGeneralSimplify rewrites dot_general(splat, X) into a reduce over the
contracting dims, scaled by the splat value and broadcast to the result shape.
Detection matched a splat operand only via m_Constant, so a
broadcast_in_dim of a rank-0 runtime scalar (e.g. a function argument)
was missed — no canonicalizer can fold it into a splat constant, so the full
dot_general survived even though it is still s * reduce(X).

Change

• Add extractBroadcastScalar, which looks through a rank-0 broadcast_in_dim
  (empty broadcast_dimensions) to its scalar operand Value. Mirrors the
  look-through in extractSplatInt, but returns a Value since the runtime
  case has no compile-time attribute.
• Detection now accepts either a splat constant or such a broadcast, on either
  operand. The constant path is unchanged; the runtime path scales the reduce
  by a broadcast of the scalar Value and always emits the multiply, relying
  on the existing downstream canonicalizer to drop it when the value is 1.
• The dot_general element-type constraint guarantees the scalar's type
  matches the reduce result, so no convert is needed.

Tests

test/lit_tests/dot_general_bcast_scalar.mlir covers a broadcast of a constant
scalar (already folded by the pipeline) plus runtime scalars on the LHS and RHS
operands, both of which now simplify away the dot_general.

Before (runtime scalar, LHS):

func.func @bcast_rank0_runtime(%s: tensor<f32>, %arg1: tensor<1024x32xf32>) -> tensor<24x32xf32> {
    %ones = stablehlo.broadcast_in_dim %s, dims = [] : (tensor<f32>) -> tensor<24x1024xf32>
    %r = stablehlo.dot_general %ones, %arg1, contracting_dims = [1] x [0], precision = [DEFAULT, DEFAULT] : (tensor<24x1024xf32>, tensor<1024x32xf32>) -> tensor<24x32xf32>
    return %r : tensor<24x32xf32>
}

After:

func.func @bcast_rank0_runtime(%arg0: tensor<f32>, %arg1: tensor<1024x32xf32>) -> tensor<24x32xf32> {
    %cst = stablehlo.constant dense<0.000000e+00> : tensor<f32>
    %0 = stablehlo.reduce(%arg1 init: %cst) applies stablehlo.add across dimensions = [0] : (tensor<1024x32xf32>, tensor<f32>) -> tensor<32xf32>
    %1 = stablehlo.broadcast_in_dim %0, dims = [1] : (tensor<32xf32>) -> tensor<24x32xf32>
    %2 = stablehlo.broadcast_in_dim %arg0, dims = [] : (tensor<f32>) -> tensor<24x32xf32>
    %3 = stablehlo.multiply %1, %2 : tensor<24x32xf32>
    return %3 : tensor<24x32xf32>
}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 31.11%. Comparing base (94c1d96) to head (8a206cb).
⚠️ Report is 2 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2637      +/-   ##
==========================================
- Coverage   31.73%   31.11%   -0.63%     
==========================================
  Files         227      227              
  Lines       44908    44875      -33     
==========================================
- Hits        14253    13963     -290     
- Misses      30655    30912     +257     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.