Batch IFD and value reads

[neeraj-pilot]

Batches contiguous IFD entries and numeric values into bounded reads, reducing async RandomAccessSource round trips without changing the API.

Across all eight bundled fixtures, the sum of median warm-cache macOS AOT parse times fell from 100.86 ms to 15.82 ms, while source read calls fell from 4,974 to 606, with exact tag and warning parity.

Fixture	Baseline	Batched	Speedup	Read calls
AVIF	7.14 ms	2.35 ms	3.0×	381 → 131
HEIC	61.97 ms	3.98 ms	15.6×	3,230 → 214
JXL	1.00 ms	0.64 ms	1.6×	43 → 24
JXL/Brotli	1.66 ms	1.55 ms	1.1×	12 → 12
PNG	2.00 ms	0.59 ms	3.4×	103 → 28
CR3	2.00 ms	1.03 ms	1.9×	19 → 19
RAF	18.75 ms	4.14 ms	4.5×	862 → 104
WebP	6.33 ms	1.55 ms	4.1×	324 → 74

The cumulative effect is:

Configuration	Summed median	Logical reads	Physical reads
Original direct source	100.86 ms	4,974	4,974
4 KiB read-ahead	8.71 ms	4,974	19
Read-ahead + this PR	5.25 ms	606	19

Together, read-ahead and batching make the original path 19.2× faster and reduce physical reads from 4,974 to 19. This PR provides a further 1.7× improvement on the read-ahead path.

[mgenware]

Published v4.2.0