Fix broadcasting gradient reduction in ADD and MUL operations

[c2akula]

Summary

Fixes gradient computation when broadcasting occurs in element-wise ADD, MUL operations and batch MATMUL operations. This resolves the TODO at src/tofu_graph.c:965 and is the first feature for v1.1.0.

Problem

When broadcasting occurs in forward pass, the backward pass was not correctly reducing gradients back to the original input shape:

1. Element-wise operations (ADD, MUL): [3,1] * [3,4] -> [3,4] - gradient had shape [3,4] but needed to be reduced to [3,1]
2. Batch MATMUL: [1,3,4] @ [2,4,5] -> [2,3,5] - gradient had shape [2,3,4] but needed to be reduced to [1,3,4]
3. Matrix transpose: For batch matmul backward pass, transpose was reversing ALL dimensions instead of only the last 2 (matrix dimensions)

Solution

New helper function: reduce_grad_for_broadcast()

• Identifies broadcast dimensions by comparing input and gradient shapes
• Sums gradients over broadcast dimensions using tofu_tensor_sumreduce()
• Handles dimension mismatches (when output has more dims than input)
• Reshapes result to match original input shape
• Manages memory properly for intermediate tensors

Updated functions:

• add_backward() - Now reduces gradients for both inputs when broadcasting occurs
• mul_backward() - Now reduces gradients for both inputs when broadcasting occurs
• matmul_backward() - Fixed to:
  • Only transpose matrix dimensions (last 2) not batch dimensions
  • Reduce gradients when batch broadcasting occurs

Testing

New Tests

Created comprehensive test suite (test/standalone/test_broadcast_gradient.c):

• ✅ MUL with simple broadcasting: [3,1] * [3,4]
• ✅ MUL with multi-dim broadcasting: [2,1,3] * [1,4,3]
• ✅ ADD with scalar broadcasting: [1] + [3,4]
• ✅ ADD with no broadcasting (baseline)
• ✅ MATMUL with batch broadcasting: [1,3,4] @ [2,4,5] -> [2,3,5]

Results: 5/5 tests pass

Existing Tests

• ✅ All unit tests pass (66/66)
• ✅ Phase 1 validation passes (gradient checking for all ops including MATMUL)
• ✅ Phase 2 validation passes (architecture tests)
• ✅ Phase 3 validation passes (multi-class classification, regression)

Impact

• Breaking: No - backward compatible
• Performance: Minimal overhead (only when broadcasting occurs)
• Correctness: Fixes incorrect gradient computation for broadcast cases across ADD, MUL, and MATMUL operations

Resolves

Part of v1.1.0 roadmap - Priority 1 (P1) bug fix