[avnoordenne]

_optimize_batch capped the memory probe with
min(batched_state.n_atoms, max_atoms_to_try). max_atoms_to_try bounds how
much InFlightAutoBatcher may try to fit on the device, so tying it to the
job's atom count made the ceiling a function of how much work was submitted.
InFlightAutoBatcher then raises when any state's metric exceeds that
ceiling. Measured on a Quadro RTX 6000 with a Lennard-Jones stand-in so only
the batching varies:

single structure               ValueError: Max metric ... 18.0 > 17.1
small then large (10, 200)     ValueError: State metric=398.0 > 302.4
3 equal conformers             ok
mixed library (12/40/8/150)    ValueError: State metric=298.0 > 211.2

All four complete after. A single structure failed deterministically: the
probe was capped at one structure's worth of atoms and the 0.95 padding put the
ceiling just under that structure's own metric. Equal-sized conformers were the
shape that survived, which is why this looked intermittent.

Also drops the batcher.load_states() pre-load. torch_sim.optimize calls it
itself on the OptimState, which carries the optimizer's own tensors and so
reflects the real footprint. The pre-load was not just redundant:
_get_first_batch skips the estimate when max_memory_scaler is already set,
so the pre-load pinned the ceiling to a measurement on the bare SimState and
torch-sim never re-measured. Worst with LBFGS, which stores history per system.

Consequence: the "Memory estimation" timing phase disappears, since the probe
now runs inside torch_sim.optimize. TimingCapture.memory_estimation is kept
so examples/full_benchmark.py still runs; it reads 0.0 and the cost lands in
"Optimization". Documented on the property.

Full suite: 244 passed, 3 skipped (no new tests; the four-shape check above was
done by hand).