Fix silently wrong targets on MPS with >= 65,536 training rows

[rminla]

Issue

Fixes #1357

Motivation and Context

On MPS with torch <= 2.14.1, F.pad(..., value=float("nan")) overwrites its input once the tensor reaches 2^16 rows, even with a zero pad width. _prepare_targets uses it on the train targets, so with >= 65,536 training rows every target came out NaN or another row's value, and predictions collapsed to near-constant with no error (details and repro in the issue).

This builds the NaN test rows with new_full and appends them with torch.cat, which gives the same result on every device. _prepare_targets is duplicated in each single-file architecture, so the change is in all five (v2, v2.5, v2.6, v3, v3.5). The comment can go once torch >= 2.15 is the minimum (fixed in nightly 2.15.0.dev20261002).

---

Public API Changes

• No Public API changes
• Yes, Public API changes (Details below)

---

How Has This Been Tested?

• New tests/test_architectures/test_prepare_targets.py: all five architectures, 65,535 / 65,536 / 70,000 train rows, with and without test rows, on CPU and (when available) MPS. On torch 2.14.1 / M4 Max: 25 MPS cases fail before the change and 60/60 pass after. CI has no MPS runner, so the MPS cases only run on Apple Silicon.
• End to end with TabPFNRegressor (v3.5 weights, device="mps", 1,000 held-out rows): MAE at 65,536 train rows 16.58 -> 7.57, at 70,000 rows 13.70 -> 7.55.

---

Checklist

• The changes have been tested locally.
• Documentation has been updated (if the public API or usage changes). Not applicable.
• A changelog entry has been added (see changelog/README.md), or "no changelog needed" label requested.
• The code follows the project's style guidelines. ruff==0.15.12 format and check show no new findings.
• I have considered the impact of these changes on the public API.

---

[CLAassistant]

All committers have signed the CLA.

[alanprior]

Thank you for the contribution @rminla, it might take us a few more days to properly review it. I'll follow up on this!