bench: add comprehensive decode benchmarks

Add decode benchmark covering 16 TIFF configurations with both a real
photo corpus (if available) and synthetic in-memory images:

Strip: uncompressed, PackBits, LZW, LZW+hpred, deflate+hpred, RGB16
deflate+hpred, gray8 LZW, gray8 deflate+hpred, RGB f32 deflate+fpred,
gray f32 deflate+fpred, palette8 LZW, bilevel fax4.
Tiled: RGB8 LZW/deflate/JPEG 256, gray f32 deflate+fpred 512.

Synthetic benchmarks isolate predictor and compression overhead with
controlled 2048x2048 images. A callgrind profiling example is included
for instruction-level analysis.

.gitignore

@@ -2,3 +2,4 @@
 /target/
 **/*.rs.bk
 Cargo.lock
+tests/benches/corpus/