Fix the two failures keeping master's CI red

[yichao-liang]

Master's CI has two distinct failures. This fixes both.

1. TypeError: cannot pickle '_abc._abc_data' object

The dominant one - it has failed whichever pytest-split shard test_main,
test_degenerate_mlp_sampler_learning, test_oracle_samplers or the GNN
suites landed on, three times in two days (runs 32418799832, 32260627910,
32258416611).

It is not collection timing, which is what #142 guessed. It is the
dill==0.3.5.1 pin. dill serialises a class by value whenever it cannot
find that exact class again at module.qualname, and before 0.3.6 the
by-value path shipped the class __dict__ verbatim - including the
_abc_impl slot ABCMeta puts on every class it builds, which holds an
unpicklable _abc_data. dill 0.3.6 added _get_typedict_abc, which strips
_abc_impl and the ABC registry caches and re-registers the subclasses on
load. Nothing else in an object graph can reach an _abc_data, so that path
is the whole bug. It reproduces directly on the pinned version and not on
0.3.9.

dill==0.3.9 pulls multiprocess 0.70.17 / pathos 0.3.3, its lockstep
releases. #142's retry is removed - it was a mitigation on a hypothesis its
own docstring said the continued failures would rule out - but the
serialise-before-writing it was built on stays, so pkl_dump_with_retry
becomes pkl_dump_all_or_nothing. Its "fake a transient failure" test is
replaced by one that pickles a genuinely by-value ABC, so a pin that slips
back below 0.3.6 fails there instead of at random in CI.

2. test_emcee_recovers_rate_params

It asserted rel_err < 0.3 on FitResult.point_estimate, which is the
single highest-log-probability draw out of 9600. The data does not identify
the rate params that sharply - water_fill_speed's 95% credible interval
runs [0.0031, 0.0277] around a true 0.0200, roughly ±70% wide - so which
draw wins the argmax moves between machines, and PyBullet trajectory
generation is platform-dependent on top of that. The same commit passed on
master's runner and failed on another at 32.2%.

Chain length is not the lever: 500 and 1000 steps return bit-identical
estimates. Percentiles of that chain are stable where the argmax is not, so
they are what gets asserted now - truth inside the 95% interval, and the
posterior median closing at least a quarter of the gap the 50% perturbation
opened.

Verification

On this branch (master + these two commits):

• 362 passed, 3 skipped across every dill-touching test file, all three GNN
  approach suites, tests/code_sim_learning/ and tests/test_main.py
• mypy: Success: no issues found in 724 source files
• yapf / isort / docformatter clean; pylint 10.00/10 on the changed files

One process note

master requires 14 checks and one review, but enforce_admins is off, so
gh pr merge --admin clears a failing check and not just the review gate.
That is how the three red merges above landed. Worth turning
enforce_admins on, or agreeing not to use --admin on anything red.