Improve numerical stability handling for near-constant features in StandardScaler

[coderabbitai]

Summary

The cuml.accel StandardScaler currently produces incorrect results when scaling near-constant features (features with very low variance). This affects numerical correctness once StandardScaler is auto-accelerated.

Problem

Test failures in test_standard_scaler_near_constant_features show that the GPU implementation lacks proper handling for features with variance close to zero (tested with values like 1e-10, 1, and 10000000000.0 at various sample sizes).

Requested Solution

Either:

• (a) Match sklearn's stability handling: Implement epsilon/variance thresholding when computing scale, matching sklearn's behavior for near-zero variance columns
• (b) Detect and force CPU fallback: Identify near-constant features during fit and dispatch to sklearn's CPU implementation (similar to other unsupported cases like partial_fit or sample_weight)

Context

This issue was identified during PR #7766 which adds cuml.accel support for StandardScaler.

• Related PR: Add cuml.accel support for StandardScaler #7766
• Discussion: Add cuml.accel support for StandardScaler #7766 (comment)
• Affected tests: 86 test cases covering dense and sparse arrays (CSR/CSC) with float32/float64 dtypes

Acceptance Criteria

• Tests in test_standard_scaler_near_constant_features pass without xfails
• Numerical results match sklearn for all near-constant feature scenarios
• Solution is consistent with other GPU limitation handling in cuml.accel

Requested by: @csadorf

[csadorf]

We should hold off working on this to see if this gets resolved with #7841 .

[jcrist]

This was resolved by #8020. Closing.