Add fast paths for FlipVertical, FlipHorizontal, and Rotate180 orientations

[hjanuschka]

Summary

Adds optimized fast paths for all 8 orientation variants in the low-memory save stage.

Non-transposing orientations:

• Identity: existing fast path (memcpy/SIMD interleave)
• FlipVertical: reuses identity with reversed output row
• FlipHorizontal: reverses pixel order within each row
• Rotate180: combines both (reversed row + reversed pixels)

Transposing orientations (input row → output column):

• Transpose: output_x = input_y, output_y = input_x
• Rotate90Cw: output_x = height-1-input_y, output_y = input_x
• Rotate90Ccw: output_x = input_y, output_y = width-1-input_x
• AntiTranspose: output_x = height-1-input_y, output_y = width-1-input_x

Supports all data formats (U8, U16/F16, F32) with 1-4 channels.

Test plan

• All 32 orientation tests pass

[github-actions]

Benchmark @ f6de453

MULTI-FILE BENCHMARK RESULTS (4 files)
  CPU architecture: x86_64
  WARNING: System appears noisy: high system load (2.10). Results may be unreliable.
Statistics:
  Confidence:               99.0%
  Max relative error:        3.0%

Comparing: 37288890 (Base) vs 0be4cd63 (PR)

File	Base (MP/s)	PR (MP/s)	Δ%
bike.jxl	21.676	21.795	+0.55% ±1.6%
green_queen_modular_e3.jxl	8.281	8.307	+0.30% ±0.4%
green_queen_vardct_e3.jxl	19.908	20.044	+0.69% ±2.3%
sunset_logo.jxl	2.325	2.333	+0.32% ±0.9%

[veluca93]

I'd prefer to just add FlipVertical for now, and think later about how to best implement other transformations.