v0.3.25

## What's Changed
* fix(verify): numerics_check compares the unrounded outputs, not the 4-dp answers by @aashish254 in https://github.com/NandhaKishorM/laya/pull/884
* fix(agent): validate CUDA ordinals, accept auto, survive capability-probe failures by @sahiixx in https://github.com/NandhaKishorM/laya/pull/886
* fix(serve): split an oversized batch across forward passes instead of collating it whole by @keemsisi in https://github.com/NandhaKishorM/laya/pull/889
* feat(calibrate): wire histogram binning into answers and the calibration payload by @Bruce-Yii in https://github.com/NandhaKishorM/laya/pull/890
* feat(ts): accept the per-bucket minConfidence map by @Bruce-Yii in https://github.com/NandhaKishorM/laya/pull/891
* fix(cli): write redirected output as utf-8 by @plox-sumit in https://github.com/NandhaKishorM/laya/pull/893
* feat(serve): add idle unload and shared-server MCP mode by @Rish-it in https://github.com/NandhaKishorM/laya/pull/894
* feat(backends): restore the inference backend class layer by @cklxx in https://github.com/NandhaKishorM/laya/pull/895
* feat(fast): add an fp32 TileLang CPU lowering for the kernels by @cklxx in https://github.com/NandhaKishorM/laya/pull/896
* fix(model): match the action head dtype so AOTInductor packaging completes by @cklxx in https://github.com/NandhaKishorM/laya/pull/897
* feat(agent): carry the compile items left out of the #472 split by @cklxx in https://github.com/NandhaKishorM/laya/pull/898
* feat(ts): port per-bucket minConfidence map to on-device engine by @Swatantra-66 in https://github.com/NandhaKishorM/laya/pull/900
* perf(fast): use two pipeline stages for GEGLU by @RagingSilence in https://github.com/NandhaKishorM/laya/pull/901

## New Contributors
* @sahiixx made their first contribution in https://github.com/NandhaKishorM/laya/pull/886

**Full Changelog**: https://github.com/NandhaKishorM/laya/compare/v0.3.24...v0.3.25