Fix snippets for autocomplete benchmark

[owtaylor]

Currently, the snippets in the autocomplete benchmark that depend on the LSP (import definitions, root path) are ending up empty.

This seems to be because the code that retrieves them from the IDE expects the autocomplete file to be the active file, which it isn't when the benchmark command is run.

[panyamkeerthana]

Running CrossCodeEval against our Granite.code and computing evaluations metrics for granite-3.3-8b-base using:

• Commented snippet template:

VS

• No snippet template:

(there's a ~2% levenshtein normalized edit similarity difference)

[panyamkeerthana]

While running CrossCodeEval with granite3.3-8b-base:

• Commented snippet template:

VS
• No snippet template:

(there's a significant difference in evaluation parameters, the commented snippet template improved the exact match by 1% but the edit similarity reduced by ~1%. The RAG snippets are retrieved based on openAI cosine similarity)