Tweak Granite template and postprocessing

Description

This PR aims to improve granite3 models tab completion support. Granite3 is not FIM trained (unlike previous versions), so needs to use the holeFillerTemplate in AutocompleteTemplate.getTemplateForModel, instead of the default stableCodeFimTemplate. Additionally, we found that smaller variants of granite3 models tend to sometimes return results that do not always satisfy the hole filler prompt instructions, typically repeating existing code in the response. Granite specific post-processing was added to postprocessCompletion to try to mitigate those misbehaviors.

Finally, I chose to remove extra lines prepended to this.writeLog(Completion:` statements, as I was having difficulty visually parsing the actual model output. Now we can really tell if extra lines were returned by the model.

Checklist

[x] No doc updates needed

Screenshots

[x] No visual changes

Testing

install granite3-dense:2b on ollama: ollama pull granite3-dense:2b

configure ~/.continue/config.json with:

"tabAutocompleteModel": {
    "model": "granite3-dense:2b",
    "provider": "ollama",
    "contextLength": 4096,
    "completionOptions": {
        "maxTokens": 2048,
        "temperature": 0,
        "topP": 0.9,
        "topK": 40,
        "presencePenalty": 0,
        "frequencyPenalty": 0.1
    },
    "title": "granite3-dense:2b"
},

Try auto completing on manual-testing-sandbox/test.js. Given:
```
multiply(number) {
this.res<cursor>
return this;
}
```
the model returns:
```
this.res *= number;
```

The end of the prefix is then removed from the completion, so it can be inserted properly

Name	Link
Latest commit	3934e8d81facccc42a44932311a4f36906aec8df
Latest deploy log	https://app.netlify.com/sites/continuedev/deploys/673f6ba24844a800089ee20f

continuedev / continue