examples: make the agent-server trial timeout configurable

[Shi-Dong]

Summary

Stacked on #1919 — base branch is shi/swe-agent-daytona-example, so review that
one first. This PR's own diff is 2 files, +24/-2.

swe_agent_function.run hardcoded a 3600 s ceiling on the agent-server /run
call, while both READMEs start the agent server with --agent-timeout 5400. The
httpx client is built with timeout=None, so that hardcoded 3600 was the
effective per-trial cap no matter what the server had been configured for — any
--agent-timeout above 3600 was dead configuration.

The inversion costs more than a mislabelled constant. For a trial that runs
between 60 and 90 minutes:

• the trainer records it as aborted and returns None;
• the agent server keeps running it, so its sandbox and one --max-concurrent
  slot stay occupied for up to another 30 minutes — on quota-limited sandbox
  backends that is exactly what pushes later trials into create failures;
• the aborted-sample dynamic filter rejects the whole GRPO group the sample
  belonged to, not just that sample. At --rollout-batch-size 4 one late trial
  removes a quarter of the step's gradient signal.

The ceiling now comes from AGENT_TRIAL_TIMEOUT and defaults to 7200, above the
5400 the examples ship. The agent server's own --agent-timeout is therefore the
one that fires first, and it tears down its own trial and releases the sandbox;
the client ceiling is only a backstop for a server that has become unreachable.
The README records that ordering requirement, which it previously contradicted by
telling users to keep --agent-timeout generous.

No new launcher flag: ExecuteTrainConfig already forwards arbitrary env vars to
the rollout actors, so the knob is reachable as

python examples/swe-agent/run.py ... --extra-env-vars 'AGENT_TRIAL_TIMEOUT=10800'

The mismatch was reported by nblintao while reviewing #1919.

Test plan

• black==24.3.0 --line-length 119 --check and isort --profile=black clean
• AGENT_TRIAL_TIMEOUT unset resolves to 7200; =10800 to 10800; =3600 to
  3600
• --extra-env-vars 'AGENT_TRIAL_TIMEOUT=10800' parses to
  {'AGENT_TRIAL_TIMEOUT': '10800'}, confirming the knob reaches the rollout
  actors without a dedicated flag
• Not exercised end to end against a live agent server: reaching the timeout
  path now requires a trial longer than two hours. The default is a strict
  relaxation of the previous hardcoded value, so existing runs can only stop
  aborting trials they used to abort.