Gate auto-inject on varying-strata + gate per-period tensors on survey

Public contract fix: when `survey_design.psu` is omitted, fit()
auto-injects psu=<group_col> with nest=False. If strata vary within
group, the synthesized PSU column reuses group labels across strata
and downstream survey resolution fails on the cross-stratum PSU
uniqueness check. fit() now detects that combination before auto-
inject runs (via the existing _strata_psu_vary_within_group helper)
and raises a targeted ValueError pointing users at the explicit
`psu=<col>, nest=True` path.

Performance fix: the cell-period allocator's per-(g,t) tensor
(U_per_period) was being allocated unconditionally in
_compute_full_per_group_contributions,
_compute_per_group_if_multi_horizon,
_compute_per_group_if_placebo_horizon, and their
_compute_cohort_recentered_inputs aggregator. That added O(G*T)
memory to the main ATT path and O(G*T*L_max) to the event-study /
placebo paths on every fit — including non-survey fits that never
consume the tensor. Adds a `compute_per_period: bool = True` param
through each helper, gates allocation + accumulation behind the
flag, and threads `compute_per_period=(_obs_survey_info is not None)`
from fit(). Downstream consumers now None-guard the per-period
recentering / _compute_se calls so the plug-in SE path runs with no
cell-level work.

Adds `test_auto_inject_with_varying_strata_raises` under
`TestSurveyWithinGroupValidation` covering the targeted error path.
Narrows the REGISTRY.md survey IF expansion Note to state the
nest=True requirement under varying strata. Softens two stale
comments (`_cohort_recenter_per_period` docstring,
`_compute_cohort_recentered_inputs` inline comment) so they match
the convention framing already in REGISTRY.md and TODO.md rather
than claiming methodological correctness.

All 352 tests pass (slow MC coverage sim included).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: igerber