feat(dppo): make the log-probability clamp configurable

[tactino]

DPPO's loss clamps every element's log-probability to [-5, 2] before forming the ratio - DPPO's reference values. This makes the bounds configuration, defaulting to exactly those values, so nothing changes until someone lifts them.

Why it matters

With dppo-policy the upper bound never binds: each denoising step's deviation is floored at sampling_noise_level, and a normal with deviation 0.1 has a log-density of at most 1.38. fpo-policy is a flow policy; its step deviation is σ_t·√dt with σ_t = level·√(t/(1−t)), which at level 0.1 runs from 0.1 at the first of its ten steps down to about 0.01 at the last. Measured on freshly initialised policies at E28's settings (Hopper, level 0.1, 4,096 samples):

	log-prob elements above 2	mean entropy
fpo-policy	50.7% - 0 at t = 1.0-0.8, then 0.37, 0.59, 0.70, 0.77, 0.83, 0.88, 0.93 at t = 0.1	−1.920
dppo-policy	0% (0.17% below −5)	0.312

The entropies are exactly what E28's tensorboards logged for these cells. A clamped element carries no gradient and contributes a ratio of exactly 1, which fits what E18 and E28 (#63) logged for fpo-policy under DPPO: approx_kl around 10⁻⁷ against dppo-policy's 10⁻⁵, clipfrac near 0, and no learning on HalfCheetah, Hopper or Walker2d - while the same policy learns all three under FPO. Whether the clamp is why is not settled here; E29 will test it against the narrow noise itself.

Change

• DPPOAlgoConfig.logprob_clamp_min = -5.0, logprob_clamp_max = 2.0; _compute_loss reads them for both the new and old log-probabilities. DPPOAlgoDistributedConfig inherits them.
• inf lifts a bound. --algo.logprob-clamp-min=-inf needs the equals form: argparse reads a bare -inf as a flag (the test pins this).

Tests

tests/test_dppo_logprob_clamp.py, written first (5 of 5 failing before):

• the defaults are the reference's bounds;
• on a fresh fpo-policy at level 0.1, 35-65% of elements are above the upper bound;
• a batch clamped everywhere gives the actor exactly zero policy gradient;
• lifting the bounds restores a larger, non-zero gradient;
• the bounds parse from the command line.

Full suite locally: 186 passed, 3 skipped, 6 xfailed, 4 xpassed.

[tactino]

E29 (#65) tested this: lifting the upper bound on E28's fpo-policy · DPPO · Hopper cell gives +7.0 / +3.0 / +7.8 against the control's +10.6 / +4.5 / +9.7 - no effect. The clamp binds as described, but it is not why fpo-policy stands still under DPPO; ten times the noise moves it (+74 to +106). So this PR stays what it is - the bounds configurable, the reference's defaults unchanged - and E29 gives no reason to change the default. E29 needs it merged to be reproducible.