Vision: Improve the bounding boxes implementation

[datumbox]

PR #2785 introduced a new utility to draw bounding boxes on images. The code changes were merged but there are a few nice-to-have improvements missing.

More specifically:

• [x] The unit-test test_draw_boxes() uses PIL instead of write_image() to save the image. We should use TorchVision's method instead.
• [x] The implementation of the draw_bounding_boxes() initializes the text font within the loop which is suboptimal. We should move it outside of the loop and initialize it once.

[oke-aditya]

Looks really minor fixes @datumbox can I take this up?

[datumbox]

@oke-aditya that would be great, looking forward to your PR. :)

[AdityaKhursale]

Hi @oke-aditya and @datumbox

Sorry I did not read that you guys have concluded assignee on this one already.
I am starting new on open source contributions. This one looked really good for the "good first issue". So, I have created PR https://github.com/pytorch/vision/pull/3075 for this one.

You guys can decide which one to merge. @datumbox really thanks for accurate labels. This really helped me to understand the flow. Also, please have a look at the PR. Your review and feedback will help me a lot.

Thank you.

[oke-aditya]

No worries @AdityaKhursale this is open source and all are welcome to do pr. You can go ahead.

[AdityaKhursale]

No worries @AdityaKhursale this is open source and all are welcome to do pr. You can go ahead.

Thank you @oke-aditya. Really appreciate.