Cast float64 targets before the device move in the context-cache path

[manjunathshiva]

Same root cause as #68, second call site: #69 fixes the float64→float32 cast ordering in _predict_step_pytorch (the uncached predict path), but _build_context_cache_pytorch has the identical pattern, and the two paths are disjoint — with cache_context=True, prediction goes through prefill/decode and never calls _predict_step_pytorch. So on Apple Silicon (device="mps"), fit(cache_context=True) still crashes after #69:

TypeError: Cannot convert a MPS Tensor to float64 dtype as the MPS framework
doesn't support float64. Please use float32 instead.

Repro (crashes at classifier_and_regressor.py:2067 on current main):

import numpy as np
from tabfm import TabFMRegressor
from tabfm.src.pytorch import tabfm_v1_0_0

model = tabfm_v1_0_0.load(model_type="regression", device="mps")
rng = np.random.default_rng(0)
X = rng.random((100, 5), dtype=np.float32)
y = rng.random(100)  # float64 — numpy's default float dtype
reg = TabFMRegressor(model=model, n_estimators=2, cache_context=True)
reg.fit(X[:80], y[:80])  # TypeError during cache build

The fix mirrors #69: do the existing conditional cast on the CPU tensor, then move. The cast stays conditional for the same reasons noted there (int64 classifier targets must keep their dtype). CPU/CUDA behavior is unchanged — those backends accept float64 and the later cast made the ordering unobservable.

The regression test fits a small regressor with cache_context=True on an MPS model with float64 targets. It fails with the TypeError above against the previous ordering, and skips where MPS is unavailable (same trade-off as #69; on CPU the old and new orderings are indistinguishable, so a CPU test would not guard anything).

Verified end-to-end on Apple Silicon (M-series, torch 2.13.0) with the real v1.0.0 checkpoints: with this fix plus #69, both TabFMClassifier and TabFMRegressor run on device="mps" with cached and uncached prediction, 100% label agreement vs CPU on breast_cancer (probability deltas ≤3e-2, consistent with bf16 kernel differences), regression deltas ~2% of target std on diabetes, and ~16–22× speedups over CPU (e.g. classifier predict 38.4s → 2.4s, cached fit 35.9s → 1.6s).