Removes dead, uncalled code from crypto/math-cuda/src/lde.rs — −256 lines, +1 (the +1 is a one-line import collapse).

What's removed

Three pub fn wrappers with no callers anywhere on current main (verified across the whole repo — the only references are their own definitions and assert_u32_domain message strings):

• coset_lde_batch_base_into_with_merkle_tree
• coset_lde_batch_ext3_into_with_merkle_tree
• coset_lde_batch_ext3_into_with_leaf_hash

They are thin non-_keep wrappers that delegate to the *_inner functions.

Interaction with #735 (why the ext3 _inner goes too)

This PR was rebased onto main after #735 ("unify & clean up the commitment layer") removed the coset_lde_batch_*_into_with_merkle_tree_keep variants. Consequences:

• coset_lde_batch_base_into_with_merkle_tree_inner — stays live, still called by coset_lde_batch_base_into_with_leaf_hash.
• coset_lde_batch_ext3_into_with_merkle_tree_inner — once refactor(stark): unify & clean up the commitment layer #735 deleted the ext3 _keep, its only remaining callers were the two ext3 wrappers above. Removing them strands it as dead code, so this PR also removes:
  • the private coset_lde_batch_ext3_into_with_merkle_tree_inner, and
  • its now-unused imports launch_keccak_ext3 / launch_keccak_ext3_row_pair (collapsing the crate::merkle import to one line).

Note: previously these wrappers were pub, which masked the dead ext3 chain (pub items are treated as reachable). Dropping the pub wrappers is what exposes the orphaned private _inner.

Verification

• Whole-repo grep on latest main: zero call sites for the three wrappers; after removal, zero references to coset_lde_batch_ext3_into_with_merkle_tree_inner and to launch_keccak_ext3*.
• Retained and still referenced: coset_lde_batch_base_into_with_merkle_tree_inner, coset_lde_batch_base_into_with_leaf_hash, GpuLdeExt3, the row-major paths (coset_lde_*_row_major_with_merkle_tree_keep), and all kept crate::merkle imports.
• rustfmt clean (the only stable-rustfmt diff is a pre-existing import-order quirk present on main, untouched here).
• Not compiled locally — math-cuda requires the CUDA toolchain (no nvcc on the dev box) and only builds under the cuda feature, which standard PR CI does not enable. A CUDA build should confirm it's warning-clean (no dead_code/unused-import) before merge.