fix(desktop): derive avatar initials by character, not code unit

[Chessing234]

Found by probing getInitials with non-Latin names; no issue filed. It feeds every avatar fallback in the app — UserAvatar, ProfileAvatar, IdentityInitialsAvatar, CommunityRail.

Two separate bugs, one commit each.

Half a surrogate pair. The function took part[0], a UTF-16 code unit. A name whose first letter lives outside the Basic Multilingual Plane is a surrogate pair, so that returned half of one — not a character:

"𠀀明"          →  "\ud840"      (a lone high surrogate, renders as �)
"𝐀da Lovelace" →  "\ud835L"

CJK Extension B appears in ordinary Chinese and Japanese given names, so every avatar for such a person rendered �. Now iterating code points, and taking two of them before joining rather than slicing two code units off the result, so the second initial can't be halved either.

A word cut in half at a combining mark. Marks are neither \p{L} nor \p{N}, so they were replaced with a separator — cutting words apart from the inside:

name	before	after
अनिल कुमार	अल	अक
नमस्ते	नत	न
မောင်မောင်	မင	မ

"अनिल कुमार" split at the vowel sign into "अन" and "ल", so both initials came from the middle of the first name and the surname was never reached; a one-word name produced two initials where there is one word to initial. Devanagari, Burmese, Thai and Khmer names take marks in ordinary spelling. \p{M} is now kept alongside letters and numbers — punctuation is still stripped, and the existing "B (relay)" → "BR" case is pinned by a test in both commits.

Verified locally at this head:

• pnpm test — 4961 passed, 0 failed (4954 before, plus the 7 new)
• pnpm check — clean; its 2 warnings and 2 infos are pre-existing and identical on main
• pnpm build — succeeded

Note: I'm an outside contributor, so the workflow runs here sit at action_required until a maintainer approves them; only the DCO check reports on its own.