Guard _fast_cache mutations; compare assert_near at narrower float dtype

[MaxGhenis]

Summary

Fixes two regressions that surfaced after 3.24.1 reached PyPI and are currently blocking policyengine-uk / policyengine-uk-data on main. Tracked in #473.

1. AttributeError: 'Microsimulation' object has no attribute '_fast_cache'

Simulation.__init__ initialises self._fast_cache = {}, but country-package subclasses (e.g. policyengine_uk.Simulation) legitimately override __init__ without calling super().__init__ — they set the attributes they need directly and don't mirror every private attribute core owns. The read-side fast path in calculate() already does getattr(self, "_fast_cache", None), but four write-side sites in set_input, delete_arrays, and purge_cache_of_invalid_values use bare self._fast_cache.pop(...) / .items() / re-assign, so they crashed on those subclasses.

Core now tolerates subclasses that don't mirror the attribute instead of requiring every downstream __init__ to set it. Regression test: tests/core/test_fast_cache_guards.py.

Repro (from policyengine-uk-data#359):

File "policyengine_uk/simulation.py", line 420, in build_from_single_year_dataset
    self.build_from_multi_year_dataset(dataset)
File "policyengine_uk/simulation.py", line 453, in build_from_multi_year_dataset
    self.set_input(variable, year, table[variable])
File "policyengine_core/simulations/simulation.py", line 1244, in set_input
    self._fast_cache.pop((variable_name, period), None)
AttributeError: 'Microsimulation' object has no attribute '_fast_cache'

2. False YAML-test failures on float-typed Variables (post-#464)

#464 switched assert_near to compare both operands at float64 so H6-style bugs (25_000_000 vs 25_000_001 silently equal under float32) would be caught. That's correct for int/float64 operands. But PolicyEngine stores float Variables as float32 (VALUE_TYPES[float]["dtype"] = numpy.float32), so YAML tests like

output:
  minimum_wage: 8.91

compare a float32-stored 8.90999985... against the literal 8.91. Under float64 comparison that diff is real (1.5e-7), so the test now fails where it silently passed before.

The float32 storage rounding isn't a regression — it's been there since forever. Make assert_near pick the comparison dtype from value.dtype:

• float32 when the value is already float32 — preserves the pre-#464 behaviour for YAML tests against float Variables
• float64 otherwise — preserves #464's H6 coverage for int/float64 operands

The existing H6 regression test (test_assert_near_detects_large_integer_difference) still passes (value here is int → promotes to float64). Added a new regression test (test_assert_near_accepts_float32_storage_rounding) asserting a float32 8.91 compares equal to the Python literal 8.91 with absolute_error_margin=0.

Downstream: unblocks the 7 pre-existing minimum_wage + UC YAML failures in policyengine-uk main and the build_from_single_year_dataset AttributeError in policyengine-uk-data.

Test plan

• 510 existing core tests still pass locally (pytest tests)
• 3 new tests in tests/core/test_fast_cache_guards.py pass
• 1 new test in tests/core/test_assert_near_precision.py passes; existing H6 test still passes
• ruff format --check + ruff check clean
• CI matrix 3.9 – 3.14 on ubuntu + windows