[GlobalOptimization] Fold empty-set reductions to their init operand. (#24949)

Reductions over a zero-sized dimension (e.g. input shape 2x0x4 -> 2x1x4)
lower to a linalg.generic with a 0-extent input, which crashes codegen
tiling (computeAllSliceParameters assertion).
    
When the reduced dimension has size 0, no input element is ever
combined,
so the result of the reduction is exactly its init operand. This adds a
pattern in RemoveZeroExtentTensors that replaces such a reduction with
its
existing init operand, so the 0-extent generic never reaches codegen.
    
Forwarding the init (instead of synthesizing a linalg.fill of the
reduction
identity) preserves the semantics of any initial accumulator value,
including
non-identity ones, and avoids materializing a new fill. For ONNX
empty-set
reductions the init already carries the identity (sum -> 0, prod -> 1),
so the observable result is unchanged.

Adds lit coverage for the rewrite:
  - a reduction with a non-identity initial accumulator (init preserved),
  - dynamic output dimensions.
    
Fixes 8 onnx_ops *_empty_set cases:
    test_reduce_sum_empty_set
    test_reduce_sum_square_empty_set
    test_reduce_sum_square_empty_set_expanded
    test_reduce_l1_empty_set
    test_reduce_l1_empty_set_expanded
    test_reduce_l2_empty_set
    test_reduce_log_sum_empty_set
    test_reduce_log_sum_exp_empty_set
    
Related to iree-org/iree#24378
    
Assisted-by: Claude Code, Doubao, DeepSeek

Signed-off-by: Yalong <yalong@multicorewareinc.com>