Add consistent NaN/Inf validation across public estimators

[TheHiddenObserver]

Context

Physical GPU validation for PR #79 passed all gates, but Gate C identified one remaining medium-severity finite-input-validation gap. The validated finite-input paths are correct; this issue tracks consistent rejection of NaN/Inf across the remaining public estimator entry points.

Goal

Introduce a shared backend-native finite-input validation contract for NumPy, CuPy, and Torch without transferring complete GPU arrays to CPU.

Scope

• identify public fit, predict, transform, and inference entry points that do not consistently reject NaN/Inf;
• validate X, y, sample_weight, cluster labels where numerical, offsets/exposure, and relevant initialization arrays;
• preserve selected backend, dtype, and device;
• avoid full-array GPU-to-CPU transfers or per-element Python loops;
• standardize exception type and actionable error messages;
• add NumPy/CuPy/Torch regression tests, including device-purity assertions;
• document any estimators that intentionally permit NaN values.

Acceptance criteria

• all public numerical entry points either reject non-finite values consistently or explicitly document supported missing-value behavior;
• metamorphic NaN/Inf tests pass across all three backends;
• device-purity audit reports no full-design transfers introduced by validation;
• complete CPU and physical-GPU suites pass.

Relationship

Follow-up to PR #79. This is non-blocking for the finite-input paths validated there.

[TheHiddenObserver]

Resolved by PR #87 and included in the 0.2.4 release tree.

Closure evidence:

• introduced shared backend-native finite validation in statgpu/backends/_validation.py;
• NumPy, CuPy, Torch, scipy/cupyx sparse arrays, pandas numerical objects, scalars, and nested sequences are covered;
• CuPy and Torch perform the finite reduction on device and synchronize only the final scalar boolean, without transferring the full design matrix to CPU;
• public estimator boundaries now guard numerical fit, predict, transform, inverse-transform, survival, scoring, metric, and inference-related entry points;
• covered inputs include X, y, sample_weight, offsets/exposure, event/time arrays, numerical clusters/groups, initialization arrays, and related public numerical arguments;
• validation failures use a consistent actionable ValueError contract and failed refits reset stale estimator/CV state;
• maintained NumPy, Torch, and CuPy regression tests include device-purity assertions;
• the complete CPU suite and exact-head physical Torch/CuPy acceptance passed.

Implementation merge: PR #87, commit 249456610c9d8ab62e9f2749e1fc7a769250b08a.
Release merge: PR #88, commit 0aeeb95b60e3e274053b8f1b6427ae50c8eec015.

The finite-input contract requested by this issue is now implemented across the current public estimator inventory.