feat: optional fast edit models

[michaelneale]

for: #2573

models that specialise editing can be used to reduce the load on the main LLM providers, increase accuracy and quality and speed and lower cost.
This supports both morphllm and relace models (and any openai compatible endpoint)

[michaelneale]

I ran it through the benchmark ("system" is default, local is with morph):

Summary of Benchmark Results

I've run the developer_search_replace benchmark using both the system goose binary and the local ./target/release/goose binary. Here are the results:

System goose Binary Results

Scores:

• 3 out of 4 runs scored 1.0 (success)
• 1 run scored 0.0 (failure)

Execution Times:

• 32.86 seconds
• 70.13 seconds
• 37.41 seconds
• 41.97 seconds

Average Execution Time: 45.59 seconds
Success Rate: 75%

Local ./target/release/goose Binary Results

Scores:

• 4 out of 4 runs scored 1.0 (success)

Execution Times:

• 50.65 seconds
• 64.52 seconds
• 75.51 seconds
• 68.33 seconds

Average Execution Time: 64.75 seconds
Success Rate: 100%

Comparison

1. Success Rate:

   • Local binary: 100% (4/4 successful runs)
   • System binary: 75% (3/4 successful runs)

2. Execution Time:

   • Local binary: Slower with an average of 64.75 seconds
   • System binary: Faster with an average of 45.59 seconds

3. Consistency:

   • Local binary: More consistent in terms of success rate (all runs succeeded)
   • System binary: Less consistent (one run failed)

The local build appears to be more reliable in completing the benchmark successfully, but it takes longer to execute. The system binary is faster on average but had one failed run.

[michaelneale]

@bhaktatejas922 I will take another look - don't worry too much about those times, the fact it is did stuff which wasn't reliably working before so measuring work vs not is tricky! we need to get better at these benchmarks

[Kvadratni]

Lgtm