perf: dynamic Dory commitment computation on the GPU should efficient…

…ly handle the offset (spaceandtimelabs#291)

# Rationale for this change
Currently the dynamic Dory commitment computation on the GPU includes
zeros for all offset values. This means adding a single commit with a
large offset will require a significant chunk of unnecessary memory.
This PR handles offsets more efficiently truncating all zero rows of the
dynamic Dory structure matrix when crating the Blitzar metadata tables.

# What changes are included in this PR?
- The `blitzar_metadata_tables` module now truncates rows of the dynamic
Dory structure matrix that are fully contained in the offset.
- Tests are updated in the `blitzar_metadata_tables` module to test the
new behavior.
- The `Gamma_2` value in the `dynamic_dory_commitment_helper_gpu` module
now uses the offset.

# Are these changes tested?
Yes

---------

Co-authored-by: Dustin Ray <[email protected]>