Validate upload caps and defer subsampled contexts to the server

[adrian-prior]

Generated by Codex

Validate training uploads against the upload cap returned by the server so valid subsampled uploads are not rejected against the per-estimator cap. Keep static shape and output checks in the client, and defer subsampled context validation and context-dependent prediction limits to the server. This avoids duplicating TabPFN's subsampling rules in the client; context errors can be reported after upload.

Validation: 20 validation, estimator, and group-column tests passed. The full configured Trunk gate passed across 57 files.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[ggprior]

Let's check if we can avoid the duplication. tabpfn-common-utils or server-side validation?
Otherwise LGTM pending this double check.

[simo-prior]

.

[simo-prior]

Thanks!