Add Fused Residual Add and LayerNorm with Saved Statistics challenge (hard #143)

[kunal-mansukhani]

New hard challenge #143: fused residual add + LayerNorm forward that also writes residual_out (x + residual) and per-row mean/rstd for the backward pass.

• Distinct from [Feat] CI workflow: Check whether submitted challenges have duplicated Questions #83 (residual + RMS norm, single output) and [feature request] ptx/sass output renders #113 (plain LayerNorm): four outputs in one pass, and large-DC-offset tests where the one-pass E[h^2]-E[h]^2 variance fails at 1e-3 (needs two-pass/Welford).
• float64-checked reference, outputs pre-filled with garbage, 12 functional tests (C=1, constant rows, large eps, offsets 500/-300, GPT-2 XL 1600, 5120 wide), perf N=16384 x C=4096.
• Starters: CUDA, CuTe, Triton, PyTorch, JAX, Mojo. validate_challenges.py: 0 errors; clang-format/black clean.

[github-actions]

✅ Review bot: low risk, safe to merge. Lint and structure checks passed; no sensitive paths or grader edits; 152 lines changed.

• ID 143 is not used by any existing challenge.
• It is distinct from #83 (residual add + RMSNorm, one output) and #113 (plain LayerNorm): it fuses the residual add with LayerNorm and writes four outputs, including saved mean/rstd. The large-offset tests also force a numerically stable variance.
• challenge.py imports only ctypes, typing, torch and the challenge base. There is no network, file, subprocess or import-time code.
• It has all 5 required methods, no init override, and name/atol/rtol/num_gpus/access_tier set. The private helpers _make and _outputs are extras, which is a minor deviation but harmless.
• reference_impl asserts shapes and dtypes with no device checks, and uses standard PyTorch ops with a float64 stable reference. The hand-computed example matches the code (rstd 1.1547 and 316.2278).
• All tensor allocations use self.device. Outputs are pre-filled with 123.0 so skipped writes fail.
• The 12 functional tests cover C=1, constant rows, large eps, zero residual, negative gamma, ragged widths, DC offsets of 500 and -300, and 1600 and 5120 widths. atol/rtol of 1e-3 is sensible and still rejects the one-pass variance on the offset tests.
• Performance test memory is about 1.07GB (four 16384x4096 float32 tensors), well under the 3.2GB limit.
• The six starters are empty stubs with only the parameter comment, and the Mojo starter uses max.gpu imports and MutAnyOrigin. challenge.html matches the signature, constraints and example.