Fix prefix removal when streaming inline completions

[krassowski]

Fixes #841

[dlqqq]

@krassowski This looks good aside from some minor typos which I've called out below, thank you for making this change! Do you recall the previous reason why we were only trimming the prefix conditionally when the response began/ended with ```? Want to verify that we're not making a regression unexpectedly.

[krassowski]

Do you recall the previous reason why we were only trimming the prefix conditionally when the response began/ended with ```? Want to verify that we're not making a regression unexpectedly.

This was just a premature optimization. Briefly, the idea was that instead of storing two versions of the suggestion text (as it is now done in this PR with suggestion and processed_suggestion) we could store only one (the processed version).

So with the previous logic (where we wanted to store only one version) if we got the following reply:

```python
print("test")
```

in four chunks:

```

python

print("test")

```

then the old logic (if we did not have the conditional) would trim the first chunk away, but accept second and third chunk (and possibly the fourth - I do not remember), so the result would be:

python
print("test")

or (if the code was only trimming the closing fence if it detected an opening fence - which I think a version of the code did)

python
print("test")
```

In either case, we would not want this to happen, the intended result is just:

print("test")

Now after this PR we are reapplying post-processing after each fragment and do not need to worry about these. While this has a negligibly higher cost, it is even more appropriate because the post_process_suggestion method is now a part of the provider API so we do not have a control on what is inside for custom providers (it may not even look at ```!).

[dlqqq]

@krassowski Thanks for lending context & addressing my feedback! Proceeding to merge.