fix(community): fix real-time streaming buffer bug for DeepInfra LLM Inference Provider

[Sumit Kumar (Sumitk99)]

Description:
In DeepInfra (langchain_community/llms/deepinfra.py), the _stream() and
_astream() methods called response.text / await response.text() before
iterating over the SSE stream. This consumed the entire HTTP response body
upfront — so all tokens were buffered in memory and delivered at once. It looked
like streaming on the surface, but every chunk arrived with Δ0.000s between
them (i.e. a bulk response in disguise).
Root cause:

# Buggy — reads the entire body, stream is now exhausted
response_text = response.text
self._handle_body_errors(response_text)
self._handle_status(response.status_code, response.text)
for line in _parse_stream(response.iter_lines()):  # nothing left to iterate
    ...

Fix:

Pass stream=True to request.post() to enable HTTP-level chunked transfer
Remove the pre-iteration response.text call; check HTTP status without reading the body
Move server-side error detection inside the per-chunk loop

# Fixed — status checked without consuming body; chunks arrive in real-time
response = request.post(url=..., data=..., stream=True)
self._handle_status(response.status_code, "")
for line in _parse_stream(response.iter_lines()):
    if line and "error" in line:
        self._handle_body_errors(line)
    ...

Issue: N/A
Dependencies: None — no new dependencies introduced.