Build the manifest-pruning predicate against the partition struct
(field.name) instead of the row schema, so it works for any partition
transform, not just identity. Falls back to a plain AlwaysTrue for
unpartitioned specs, and combines with existing predicate-based pruning
via delete_by_predicate.

…field

_OverwriteFiles deletes by exact DataFile identity, so an end-to-end
delete would still pass even if pruning silently degraded to a
non-discriminating fallback. This test inspects the built predicate
directly to guard the partition-domain fix itself.

- Comment explains that the AlwaysTrue fallback only disables the
  manifest-pruning optimization; deletion still happens by exact
  DataFile identity in _OverwriteFiles, so no rows are unexpectedly
  dropped.
- Test now asserts on the file-path set before/after deletion (exact
  path removed, count drops by exactly one) instead of relying on
  row-count alone.

…ity-manifest-pruning

- Fix stale comment in _manifests(): the call no longer touches
  self._predicate, only self._delete_files_partition_filters.
- Extract build_field_value_predicate/build_records_predicate into
  pyiceberg.expressions, shared by Transaction._build_partition_predicate
  and _build_delete_files_partition_predicate, removing near-duplicate
  EqualTo/IsNull/And/Or construction.
- Reset self._delete_files_partition_filters at the top of
  _build_delete_files_partition_predicate so it starts clean on every
  _manifests() pass, including retries.
- Update stale test_commit_retry.py docstring: conflict detection for
  the CoW-rewrite path now uses the user's delete filter directly
  (self._predicate is no longer widened by the deleted-files partition
  predicate), matching Java's approach and fixing a latent inconsistency
  where _predicate accumulated across retries while partition_filters
  stayed frozen from the first attempt. This is a behavior change beyond
  the original bug fix and should be called out in the PR description.

- Rename tests/table/test_delete_data_file_manifest_pruning_bug.py to
  test_snapshot_manifest_pruning.py, matching the repo's component-named
  test file convention (no _bug suffix).
- Drop the f"..._{catalog.name}" identifier suffix: all three catalog
  fixture params share name="test_catalog", so it was a no-op: isolation
  already comes from the per-test tmp_path.
- Add test_delete_data_file_manifest_pruning_bucket_on_same_result_type_succeeds:
  a BucketTransform over an IntegerType source column, where the pre-fix
  predicate's type happened to match the source column's type. Unlike the
  string-source case (a loud TypeError), this variant let the buggy
  row-domain predicate bind successfully and silently prune away the
  manifest containing the target file, so delete_data_file reported
  success without deleting anything. Verified this fails (silently, no
  exception) against the pre-fix code and passes with the current fix.

…ing.py

They build predicates over Record values, not expression AST nodes, so
pyiceberg/expressions/__init__.py (which only defines the expression
classes themselves) was the wrong home. pyiceberg/partitioning.py already
owns Record/PartitionSpec semantics and both call sites already import
from it, so this is a natural fit with no new import needed and no
circular-import risk (partitioning.py -> expressions is a new but
one-directional edge; expressions/__init__.py does not import
partitioning).