fix: stop forecast() from mutating the caller's input list

[ihubanov]

TimesFM_2p5.forecast mutates the list the caller passes in. Unrelated to #465/#466/#467, which only touch examples; this is in the main inference entry point.

The bug

When the batch needs padding, src/timesfm/timesfm_2p5/timesfm_2p5_base.py:168 does:

inputs += [np.array([0.0] * 3)] * (self.global_batch_size - w)

+= on a list is an in-place extend, so the padding is appended to the caller's own list:

>>> mine = [series, series, series]        # global_batch_size = 4
>>> len(mine)
3
>>> model.forecast(horizon=8, inputs=mine)
>>> len(mine)
4
>>> len(mine[-1])
3

The first call still returns the right answer, because the output is trimmed to num_inputs. The problem is the second one. num_inputs is recomputed from the now-longer list, so the caller gets back a fourth forecast for a length-3 padding array, presented as a real result, and the list grows again on every subsequent call. Silently wrong output rather than an exception.

I hit this for real while backtesting: 456 context windows became 464, with eight length-3 zero arrays appended, and the next np.array(ctxs) raised inhomogeneous shape. That failure was loud only because I happened to rebuild an array from the list afterwards — code that just calls forecast in a loop gets no signal at all.

The fix

Build the padded batch as a copy:

inputs = list(inputs) + [np.array([0.0] * 3)] * (self.global_batch_size - w)

One line, no behaviour change for a single call. As a side effect any sequence now works, where a tuple previously raised TypeError: can only concatenate tuple (not "list") to tuple on the +=.

Tests

tests/test_forecast_inputs.py covers the padded path, the evenly-divided path (no padding, no mutation), repeated calls on one list, and tuple input. They drive the base class through a stub decode function, so they run without model weights and add no network dependency to CI.

Verified they actually catch the bug: 4 of the 5 fail on master and all pass with the fix. The one that passes either way is the evenly-divided case, which never enters the padding branch.

$ pytest tests/ -q --ignore=tests/test_model_loading.py
64 passed

test_model_loading.py is excluded only because it needs the flax extra, which is unrelated to this change.