fix(deepgram): include word confidence for stt v1 alternatives

[Snowfly0709]

Closes #7180 .

What

Forward Deepgram's per-word confidence into TimedString at the two
places the plugin builds them.

Why

This is #5034's fix applied to the other module. stt_v2.py forwards
Deepgram's per-word confidence into TimedString; stt.py builds the same
TimedString at two sites without it, so words from the v1 STT always
report NOT_GIVEN. As #5034 put it for v2, the confidence "wasn't set on
the word level, even though it is used to set the overall combined
confidence" — v1 likewise reads alt["confidence"] for the utterance and
drops the per-word values.

The field is declared:

class TimedString(str):
    confidence: NotGivenOr[float]   # NOT_GIVEN when unavailable

and Deepgram returns confidence on every entry of words[].
livekit-plugins-assemblyai and livekit-plugins-google populate it too,
so v1 is now the outlier.

The raw JSON is discarded inside the plugin, so downstream code cannot
recover the value: Agent.stt_node receives already-parsed TimedStrings,
and UserInputTranscribedEvent carries only transcript: str.

Per-word confidence lets an agent tell which part of a spoken value it is
unsure about, rather than only that the utterance might be wrong — which
is what matters for structured values like email addresses and reference
numbers.

Default value

word.get("confidence", NOT_GIVEN) rather than 0, matching the field's
declared NotGivenOr[float] contract: a consumer ranking words by
confidence needs "unknown" to stay distinguishable from "zero confidence",
since those call for opposite handling.

Changes

Two lines, no new imports (NOT_GIVEN is already imported at the top of
the file):

• live_transcription_to_speech_data
• prerecorded_transcription_to_speech_event

Verification

ruff check and ruff format clean under the repo's own config. Both
functions are pure
dict -> SpeechData, so this is checkable without a network call:

from livekit.agents.utils import is_given
from livekit.plugins.deepgram.stt import prerecorded_transcription_to_speech_event

event = prerecorded_transcription_to_speech_event(
    "en-US",
    {
        "metadata": {"request_id": "r"},
        "results": {
            "channels": [
                {
                    "alternatives": [
                        {
                            "transcript": "hello world",
                            "confidence": 0.98,
                            "words": [
                                {"word": "hello", "start": 0.0, "end": 0.4,
                                 "confidence": 0.99},
                                {"word": "world", "start": 0.4, "end": 0.9,
                                 "confidence": 0.42},
                            ],
                        }
                    ]
                }
            ]
        },
    },
)

words = event.alternatives[0].words
assert [w.confidence for w in words] == [0.99, 0.42]   # before: [NOT_GIVEN, NOT_GIVEN]
assert all(is_given(w.confidence) for w in words)

I did not add a test file — tests/ has no unit tests for this plugin's
parsing helpers, so I did not want to establish a pattern uninvited. Happy
to add one wherever you'd like it.

Behaviour change

Additive, and the same shape as #5034. Words gain a value where they
previously had NOT_GIVEN; nothing that already worked changes. Consumers
gating on is_given() keep working either way.

[CLAassistant]

All committers have signed the CLA.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.