[FlagGems Operator Development Competition] Add roll operator

[HyosungSink]

Summary

Add roll operator for the FlagGems Operator Development Competition.

This PR adds:

• torch.roll(input, shifts, dims=None)

Implementation

• Added optimized Triton implementation in src/flag_gems/ops/roll.py
• Added roll correctness coverage in tests/test_unary_pointwise_ops.py
• Added benchmark coverage using the FlagGems built-in benchmark framework
• Added operator-specific benchmark shapes in benchmark/core_shapes.yaml

Implementation details:

• flatten fast path for dims=None
• specialized 1D / first-dim / last-dim kernels for hot single-dimension cases
• fused gather kernel for contiguous hot multi-dimension cases
• PyTorch-compatible argument validation, repeated-dim semantics, and empty-tensor out-of-range dims handling
• unsupported fast-path dtypes and layouts fall back safely to PyTorch behavior

Accuracy Validation

Correctness tests cover:

• dims=None flatten behavior
• single-dimension and multi-dimension roll
• positive / negative shifts and dims
• repeated dims behavior aligned with PyTorch
• scalar tensors and empty tensors
• bool tensors and float64 fallback
• non-contiguous input
• invalid shift/dim combinations and out-of-range dims

All outputs are compared against PyTorch reference behavior.

Note: the upstream roll test report contains 51 skipped cases. These are expected parameterized skips for invalid shape/dim combinations in the shared POINTWISE_SHAPES set, not hidden failures.

Benchmark

Validation results:

• pre-commit: passed
• ARCH=7 python -m pytest tests/test_unary_pointwise_ops.py -m roll -q: 257 passed, 51 skipped, 3397 deselected
• official roll benchmark harness: passed

Official benchmark harness

DType	Minimum Speedup
float32	1.080x
float16	1.250x

Representative benchmark cases

DType	Case	Speedup
float32	flat_1d_small	1.097x
float32	flat_3d_medium	2.419x
float32	single_dim_2d_last	1.814x
float32	single_dim_3d_first	2.931x
float32	single_dim_3d_last_large	2.402x
float32	multi_dim_3d	4.105x
float32	multi_dim_4d	4.886x
float16	flat_1d_small	1.103x
float16	flat_3d_medium	2.709x
float16	single_dim_2d_last	1.691x
float16	single_dim_3d_first	5.599x
float16	single_dim_3d_last_large	2.316x
float16	multi_dim_3d	4.502x
float16	multi_dim_4d	4.691x

Test Coverage Checklist

• API branches: torch.roll(input, shifts, dims=None)
• Input dims: flatten / single-dim / multi-dim
• Input scales: small / representative / hot large cases
• Tensor layout: contiguous and non-contiguous
• Tensor shapes: scalar, empty, regular multi-dim tensors
• DTypes: float16, float32, bool, float64 fallback
• Error behavior: invalid shift/dim combinations and out-of-range dims

[CLAassistant]

All committers have signed the CLA.

[HyosungSink]

Hi @douxetpur, the branch has been updated to the latest master and merge conflicts are resolved. Could you please help approve the pending workflows when convenient? Thanks!

[HyosungSink]

Hi @douxetpur, I have fixed the issues from the previous failed checks and verified the formatting fix by passing code-style in my fork-side precheck PR. Could you please help trigger a fresh round of CI/checks for this PR when convenient? Thank you!

[HyosungSink]

Hi @douxetpur, the latest python-op failure seems to be caused by a transient runner issue during actions/checkout@v6 (Failed to connect to github.com port 443: Connection timed out), before the actual test script started.

Could you please help rerun the CI when convenient? Thank you very much.

[HyosungSink]

All required checks are passing now, and the branch is up to date with the latest fixes. Could a code owner please review when convenient? Thank you very much.

[douxetpur]

LGTM

[HyosungSink]

Hi maintainers, thank you very much for your time and effort in reviewing and merging this PR.

My name is Zhiyuan Chen (陈致远). If any further follow-up or coordination is needed regarding the competition submission, I will respond promptly and cooperate actively.

Thank you again for your support.