Add the ability to translate AI image descriptions

[seanbudd]

Is your feature request related to a problem? Please describe.

Currently AI image description are only in English.
An automated translations tool is required.

The ability to translate image descriptions requires some form of model manager to install AI translations tools.
This means this work is blocked on #18662 and #19337.

Describe the solution you'd like

Update the model manager proposed in #19337 to allow installing translation models.
Include the download size, input languages, and output languages.
NVDA should automatically translate AI image descriptions using a model that supports English to the set language.

As part of this also consider adding the ability to translate any text via this models while reading.

Describe alternatives you've considered

No response

Additional context

No response

[cary-rowen]

I only have one point to make: It is crucial to select a translation model that supports multiple languages.

Unfortunately, I haven't found any small, satisfactory offline translation models yet, especially for bidirectional Chinese-English translation. If I had to name one that is usable, the offline model built into Chrome 138 seems to be the most ideal I've encountered so far. The next best might be a Gemma model of at least 4B or larger.

However, my real concern is this:
Based on #19298, the current LLM descriptions have a fairly severe hallucination problem, leading to a general consensus that the image recognition results are not very useful. I'm genuinely worried that introducing yet another translation model with subpar performance will further amplify this poor experience and increase user frustration.

[makhlwf]

I suggest using these

https://github.com/mozilla/translations
https://github.com/mozilla/firefox-translations-models

[seanbudd]

@cary-rowen @makhlwf - thanks for your research here.
With #19337, we want to be able to support a range of models that directly provide image description in local languages, and a variety of translation models to cover all languages NVDA supports.

Based on #19298, the current LLM descriptions have a fairly severe hallucination problem, leading to a general consensus that the image recognition results are not very useful. I'm genuinely worried that introducing yet another translation model with subpar performance will further amplify this poor experience and increase user frustration.

As somewhat mentioned here and on #19298, translation work is blocked on #19337 and #18662, which should result in providing higher quality descriptions