fix: extract text from tiptap bodies, and read the right description field

[ccleberg]

Checking the library against the live API disproved what its own comments claimed
— including the ones I wrote in #2. Comment and Description hand back raw
JSON blobs, and the parser this package has always carried could not have
prevented it.

DeviantArt does not use Draft.js anymore

It uses tiptap:

Draft.js, what we parse:  {"blocks":[{"text":"..."}]}
tiptap, what DA sends:    {"version":1,"document":{"type":"doc","content":[...]}}

Across 34 live comments on one deviation: 33 tiptap, 0 Draft.js, and 33
whose .Comment was the raw JSON. The old parser matched the {...} shape,
looked for "blocks", found none, and returned its input unchanged.

flattenMarkup now walks the tiptap tree — text nodes joined per block,
hardBreak as a newline, HTML entities decoded — and keeps the Draft.js path for
older bodies that still carry it. Nodes with no text of their own (images,
galleries, emotes) drop out, as they always did.

The description had a second, larger bug behind that one

It lives in Extended.DescriptionText, not the TextContent that GetDeviation
read. Of 24 deviations sampled, 23 populated the former and 1 the latter — so
Post.Description was empty for nearly every deviation, regardless of parsing.
The txt[1] == '{' guard removed in 3dad0ea was never what stood in the way; I
misdiagnosed that, and the "behavior change" I warned about in #2 never happened.
Now prefers Extended.DescriptionText and falls back, since either may be set.

Verified against the live API, not by reading

	before	after
comments whose .Comment is raw JSON	33/34	0/34
descriptions returning ""	23/24	0/24

Entities decode (I’d → I’d), and the "version":"1" string case — DA
sends version as both a number and a string — parses. That one matters: any
typed field for version fails to unmarshal on a fifth of bodies, so it is
deliberately not modelled.

Docs

Corrects the comments that asserted the Draft.js story on pkg.go.dev, and records
that flattening drops images, links, and emotes — read TextContent for those.

18 tests pass; go vet and golangci-lint clean.