fix(aggregation): Apply ExcessMTL exponentiation in log domain

[SajalDevX]

Problem

If a task has a zero gradient at the call that sets its baseline excess risk (a head that is still frozen, a loss that is not active yet, an all-zero first batch), ExcessMTLWeighting never recovers:

>>> a = ExcessMTL()
>>> J = torch.randn(2, 5); J[1] = 0
>>> a(J)
tensor([ 0.5143, -1.3016,  0.1904,  0.5383, -2.2004])
>>> a(torch.randn(2, 5))
tensor([nan, nan, nan, nan, nan])       # and every call after this

_initial_w[1] is 0, so from the next call on w[1] / (0 + 1e-7) is ~1e7, torch.exp(w * eta) overflows to inf, and weights / weights.sum() is inf / inf = nan. The same happens for an all-zero matrix at the first call.

Fix

Do the exponentiated gradient update in log space:

weights = torch.softmax(torch.log(weights) + w * self._robust_step_size, dim=0)

This is the same update (weights * exp(eta * w), normalised) whenever the old one was finite — test_log_space_update_matches_direct_formula checks that against the direct formula — but a huge excess risk now saturates the weights (that task → 1, the others → 0) instead of overflowing. It does not change the baseline handling, so behaviour stays aligned with the official implementation / LibMTL apart from no longer producing nan.

Tests

• test_zero_baseline_excess_risk_keeps_weights_finite and test_all_zero_matrix_at_baseline_keeps_weights_finite: both fail on main with nan.
• test_log_space_update_matches_direct_formula: equivalence on regular inputs.
• tests/unit/aggregation: 963 passed. Changelog entry added under Unreleased / Fixed.

[ValerianRey]

cc @KhusPatel4450