Increase tool call read timeout 30s -> 120s and add exponential backo…

[earakely-scale]

I ran the MCP atlas eval like so:

uv run python mcp_completion_script.py \
  --model "openai/gpt-4o" \
  --input_huggingface "bhertz/mcp_dummy" \  # TODO switch to dataset on ScaleAI
  --output "mcp_eval_4o_results.csv"

and experienced many MCP read timeouts https://pastebin.com/z6MH1K1E .

The current behavior does a 10s delay prior to retry causing possible thundering herd behavior. This can be partially mitigated using exponential backoff + jitter. I also increased the read timeout from 30s to 120s since I still saw some MCP tools consistently timeout (e.g. OSM); this is likely due to these tools simply having a larger response time.

After this change I found that nearly all read timeouts were mitigated; less than a handful occurring afterwards. I still think there is room to improve by having a granular tool-dependent timeout (e.g. some tools take longer than others and thus should have larger timeouts).