[Release/2.13][BMG] 59 UT new failures vs release/2.12

## Summary

Release/2.13 vs release/2.12 BMG comparison shows 59 UT new failures (baseline PASSED/SKIPPED → target FAILED/ERROR).

## Failure Breakdown

| Category | Count | Tests | Related Issue |
|----------|-------|-------|---------------|
| test_decomp.py (adaptive_max_pool) | 12 | test_comprehensive_nn_functional_adaptive_max_pool{1d,2d,3d}_xpu_{bfloat16,float16,float32,float64} | NEW |
| test_ops.py (fake_crossref fft/unpool) | 10 | test_fake_crossref_backward_{amp,no_amp}_fft_{hfft2,hfftn,irfft2,irfftn}_xpu_float32, test_fake_crossref_backward_{amp,no_amp}_nn_functional_max_unpool2d{,_grad}_xpu_float32 | NEW |
| test_linalg.py (lu_family) | 4 | test_linalg_lu_family_xpu_{complex128,complex64,float32,float64} | #3833 (closed, regressed) |
| test_linalg.py (lobpcg/svd/pca) | 4 | test_lobpcg_ortho_xpu_float64, test_pca_lowrank_xpu, test_svd_lowrank_xpu_{complex128,float64} | NEW |
| test_sparse.py (binary_operation_mul) | 4 | test_binary_operation_mul_Sparse{BSC,CSC}_xpu_{complex128,float64} | #2209 (closed, regressed) |
| test_sparse.py (index_select) | 2 | test_index_select_empty_and_non_contiguous_index_xpu_float64, test_index_select_exhaustive_index_large_xpu_float64 | NEW |
| test_foreach.py (xpu_graph) | 4 | test_big_num_tensors__foreach_norm_use_xpu_graph_True_w_empty_{False,True}_xpu_{float32,float64} | NEW |
| export/test_hop.py (invoke_quant) | 4 | test_{aot_export,pre_dispatch_export,retrace_export,serialize_export}_invoke_quant_simple_xpu_float32 | NEW |
| functorch/test_control_flow.py | 4 | test_while_loop_op_int_carry_compile_dynamic_True_backend_eager, test_{function_schema_gen,list_gen_schema_type_ScriptObj,type_gen_schema_type_ScriptObj} | NEW |
| test_fake_tensor.py | 3 | test_fake_tensor_prefer_device_type, test_add_one_dim_single_elem_cpu_with_cuda_tensor, test_op_with_zero_dim_bypassed | NEW |
| test_ops_gradients.py (stft) | 2 | test_fn_grad_stft_xpu_complex128, test_fn_gradgrad_stft_xpu_complex128 | NEW |
| test_ops_fwd_gradients.py (stft) | 1 | test_fn_fwgrad_bwgrad_stft_xpu_complex128 | NEW |
| test_ops.py (arange) | 1 | test_errors_arange_xpu | NEW |
| test_nn.py (unflatten) | 1 | test_unflatten_invalid_arg | #3930 (closed), #3907 (open) |
| test_torch.py (unflatten) | 1 | test_unflatten | #3930 (closed), #3907 (open) |

## Environment
- **Platform:** BMG (Battlemage)
- **Target:** release/2.13
- **Baseline:** release/2.12
- **Workflow runs:** target [27456522400](https://github.com/intel/torch-xpu-ops/actions/runs/27456522400), baseline [27490566817](https://github.com/intel/torch-xpu-ops/actions/runs/27490566817)

## Related Issues
- #3833 — linalg_lu_family regression (closed, now regressed again)
- #2209 — sparse binary_operation_mul (closed, now regressed again)
- #3907 — test_unflatten (open, skip issue)
- #3930 — test_unflatten named tensor removal (closed)

### Sub-issues (tracked from this parent issue)

- [ ] **#4100** — foreach_norm xpu_graph segfault
- [ ] **#4101** — FakeTensor error type mismatch
- [ ] **#4102** — stft device_lost in gradcheck
- [ ] **#4103** — while_loop uses aten._assert_scalar
- [ ] **#4104** — invoke_quant export SpecViolationError
- [ ] **#4105** — HopSchema libtorchbind_test.so not found
- [ ] **#4106** — arange wrong error for dtype mismatch
- [ ] **#4107** — adaptive_max_pool decomp output length mismatch
- [ ] **#4108** — fake_crossref fft stride mismatch
- [ ] **#4109** — linalg lobpcg/svd_lowrank timeout

Start with the listed Python test files and compare target workflow run 27456522400 with baseline run 27490566817 to reproduce the 59 BMG failures. Use sub-issues #4100–#4109 to follow each failure category; done means the target-only failures are explained and addressed or tracked there.

Written by the indexing model from the issue text.