[Integrate] Update torch-mlir to 528d3dcb (#24908)

Updates torch-mlir from its July 7 upstream baseline to
[`llvm/torch-mlir@528d3dcb`](https://github.com/llvm/torch-mlir/commit/528d3dcb9a2288743b62df9f4937b1689787e38a),
using
[`iree-org/torch-mlir@137b8fb8`](https://github.com/iree-org/torch-mlir/commit/137b8fb8596f4c8091f69535e5153f35b7ece311)
(`sm-iree-integrates/torch-20261002`) adapted to IREE's current LLVM
revision.

On top of canonical torch-mlir, the integration branch carries the
adaptation patches already on IREE's current pin that upstream does not
have yet: the LLVM submodule bump, `SymbolOpInterface` for
`ClassTypeOp`/`GlobalSlotOp`, and the TMTensor strict-properties
assembly fix. The `DenseMapInfo` and `RegionSuccessor` adaptations from
the previous pin are now upstream and were dropped.

ONNX op test expectations:

- Now passing, removed from `expected_run_failures` in the CI-exercised
configs:
`averagepool_3d_dilations_large_count_include_pad_is_0_ceil_mode_is_True`,
`nonmaxsuppression_flipped_coordinates`, `reduce_{max,min}_bool_inputs`.


- `test_reversesequence_time` now fails to compile on llvm-cpu and is
marked as an expected compile failure. ReverseSequence now lowers to an
`scf.for` over the time axis, so each dispatch reverses a column picked
by a dynamic offset. The resulting `vector.gather` reads a strided
column subview with a dynamic offset, which neither
`RemoveStrideFromGatherSource` (bails on dynamic offsets) nor
`Gather1DToConditionalLoads` (rejects non-unit inner stride) handles, so
the gather reaches LLVM conversion unlowered. The proper fix is
dynamic-offset support in the upstream MLIR gather lowering.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Signed-off-by: Thomas Ziereis <ziereis@roofline.ai>