docs: fix flag help and fault-tolerance default values

[YeonwooSung]

Align help strings and published docs with the current code. No runtime behavior changes.

Fault-tolerance docs disagree with code defaults

docs/en/advanced/fault-tolerance.md and docs/zh/advanced/fault-tolerance.md listed stale health-check defaults.

Flag	Before (docs)	After (matches slime/utils/arguments.py)
--rollout-health-check-first-wait	300	0
--rollout-health-check-interval	10	30
--rollout-health-check-timeout	5	30

The example still sets --rollout-health-check-first-wait 600 as an override for large-model kernel compilation, not as the default.

--disable-rollout-global-dataset help is inverted

action="store_false" / dest="rollout_global_dataset" means passing the flag disables the global dataset.

• Before: "If set, the rollout will use --prompt-data... If not set, you need to manage the data by your self."
• After: Default uses --prompt-data as a global prompt dataset. If this flag is set, you manage data yourself (no global dataset).

--kl-loss-type help advertises invalid kl

• Before: kl, k2, k3, low_var_kl
• After: k1, k2, k3, low_var_kl (default k1), matching choices.

CONTRIBUTING.md published doc URLs 404

Fixed both EN and ZH sections:

• speculative_decoding.html → https://thudm.github.io/slime/en/advanced/speculative-decoding.html
• low_precision_training.html → https://thudm.github.io/slime/en/advanced/low-precision.html
• reproducibility.html was already correct