1. It's Turing-complete, which means you can just use any other language for XML manipulation as well such as JavaScript (has DOM language bindings) or Prolog (rule/pattern based like XSLT) or maybe LISP (like DSSSL was). The benefit is that using a mainstream language gives you better mindshare and infrastructure such as APIs for DB access, IDE support for testing, etc. which I found inevitably necessary in every large XSLT project
2. XSLT 2 and 3 only has a single implementation (Saxon) by the same author as the XSLT spec itself - not my idea of a standard (and, in fact, not meeting W3C's criteria of at least two interoperable implementations).
XSLT works well if you need lots of literal XML content to be produced to the output, but will get verbose, degenerative, and awkward (priority rules) for highly dynamic content.
I don't think I agree about responsive css, but that's another discussion.
It was originally written in XSLT 1.0 so that it could generate in the web browser, which had some serious limitations and was incredibly verbose. XSLT 2 is much less verbose, but obviously XML is verbose in general.
If I had to do it over, I'd use Lisp, but the functional and template-based style of XSLT made for very natural syntax-directed translation for the compilers. I still have the cruft of old XSLT 1 code (which was written when I was also _learning_ XSLT), but all things considered, it's not all that bad. Most of the pain points are the patchwork evolution of the project; done from scratch in XSLT, it'd be much more elegant.
And, like Lisp, XSLT can parse itself as data (sans xpath). That allowed me to use a semi-literate style (generate literate documentation from the sources), for example.
[0]: https://github.com/lovullo/tame [1]: https://github.com/lovullo/liza-proguic/tree/master/src [2]: https://github.com/lovullo/literate-xsl with example output https://mikegerwitz.com/hoxsl/manual/
It's not without its wobbly bits, but show me a language that isn't.