feat: optimize proof submission

[alexghr]

Further optimize the proof submission flow:

1. hash haeader from calldata instead of memory (avoids a copy)
2. read header hashes in bulk (avoid checks run on every iteration of a loop)
3. reuse already read header hashes

Other than submitting a proof of the first checkpoint (which is +2k gas compared to next) every other scenario is cheaper than the code on next.

Scenario	Current gas	vs #25404	vs origin/next
1 fresh	663,452	−1,506 (−0.23%)	+2,243 (+0.34%)
8 fresh	965,328	−21,594 (−2.19%)	−14,985 (−1.53%)
8→16 extension	930,160	−28,490 (−2.97%)	−57,867 (−5.86%)
16 fresh	1,256,565	−44,700 (−3.44%)	−34,869 (−2.70%)
32 fresh	1,729,965	−91,388 (−5.02%)	−75,071 (−4.16%)