feat(slime): multi-turn trace merging, validation at an input dataset and gateway config fix in custom rollout function

[lyzustc]

This PR adds drift-free multi-turn trajectory merging to the slime custom rollout
function and fixes several issues that surfaced when running on full training /
validation datasets (eval episodes timing out and crashing the engine, connection-pool
churn, and unreadably noisy logs). It also brings SlimeRunner and the example scripts
in line with these changes.

Custom rollout function (integration/)

• feat — auto-merge multi-turn conversations (traces.py). Added
  merge_traces_to_samples: an episode's turns are folded into one training
  Sample per contiguous prefix-extending segment (bridge tokens masked out,
  completions trained), instead of one Sample per turn. Pairs with the gateway's
  cumulative token mode; falls back to one-Sample-per-turn when a turn breaks the
  prefix.
• feat — cumulative token mode plumbing (rollout.py, gateway.py). New
  cumulative_token_mode / renderer_family config knobs, forwarded to the
  gateway; merge_traces_to_samples is wired in as the rollout's sample builder.
• feat — validation (rollout.py). Supported doing validation on an input validation
  set periodically with an interval in training. Bounded in-flight agent sessions in validation
  by max_concurrent in yaml config.
• fix — connection-pool churn (rollout.py). Added max_pool_connections
  so user can set >= max_concurrentto stop boto3 client logging "Connection pool
  is full, discarding connection" under concurrent ACR sessions.
• fix — log noise (rollout.py, gateway.py). Replaced per-session logging
  with a single per-batch summary (episodes / succeeded / failed / sequences);
  failed episodes log once at INFO with the full ACR session_id for CloudWatch
  lookup. Added gateway_log_level (default warning) to silence the gateway's
  uvicorn/httpx access logs.
• fix — gateway always uses SGLang /generate (gateway.py). The slime
  backend now enables use_sglang unconditionally and always passes the served
  --model; removed the dead use_sglang config field. (The /generate
  mechanics live in the rllm-model-gateway PR.)

Training scripts & runner

• runner.py — exposed the new knobs on SlimeRunner
  (sglang_tool_call_parser, sglang_reasoning_parser, cumulative_token_mode,
  renderer_family, max_pool_connections, gateway_log_level,
  sglang_context_length) and added cuda_home to pin CUDA_HOME/LD_LIBRARY_PATH
  (incl. --train-env-vars for the Megatron actors), fixing TransformerEngine's
  "Multiple libcudart" abort. The tool-call parser is now a field instead of a
  hardcoded qwen25.
• examples/math_agent/train.sh — ported the CUDA-toolchain pinning,
  torch_memory_saver fixup, eval flags, log suppression, and checkpointing from
  the working run; all paths/credentials stay env-var placeholders.
• config.yaml.example — documented the new max_pool_connections,
  cumulative_token_mode, and renderer_family settings.

This PR must work together with rllm-org/rllm#715.
A follow-up PR with slime environment setup guide in docs changes will come soon.