More configurable timeout values in config.toml

[evilJazz]

Following my earlier PR #15 here is another change that removes more hard-coded timeout values from the code and moves them into the config.toml.
I also de-duplicated the duration parsing logic by moving it into a new reflection-driven helper function parseDurations that reads the existing toml and newly added default struct tags. All timeouts, intervals and durations have been moved into a composite struct ProxyConfig.Durations and the code has been refactored accordingly.

This fixes an issue in conjunction with my local llama-cpp instance when running very long LLM sessions where creating an answer by the LLM might take over 10 minutes (600 seconds). llama-cpp would always show "cancel task" in the log files, causing the output to be truncated and agent to re-iterate over and over again (and the advisor-LLM in omp to get "angry" LOL).

Thank you for this project! You are saving a lot of power keeping my space-heater server "Deepthought" powered down most of the time. :D

[darksworm]

Thanks for the PR — making these timeouts configurable is the right call, and the defaults matching the old hard-coded values keeps existing setups safe. A few asks before merging:

1. Replace the reflection-based parseDurations with plain explicit parsing (or a small parse(name, value, default) helper). The match-by-field-name contract between Config and Durations fails silently — a misspelled field name drops a duration to zero with no compile error and no test catching it. Explicit code is about as short and much safer here.

2. Formatting: the reflect import is unsorted (before crypto/tls), and the continuation line in the waitForWake timeout error lost its indentation. gofmt/goimports will fix both.

3. Document that "0" disables a timeout in the README — that's what you'll want for your >10-minute LLM responses, since the 10m default will still cut them off.

Happy to merge after that. Thanks again!