fix: preserve the SNIPSTART marker's indentation when splicing

[lennessyy]

What was changed

spliceFile wrote snippet.fmt() output verbatim, always starting at column 0. A SNIPSTART marker does not have to be at column 0 — in Markdown it can be nested inside a list item, and in MDX inside a JSX element. Writing at column 0 under an indented marker ends the enclosing block.

This re-applies the marker's own leading whitespace to every non-blank line written between the markers.

Why

Two failure modes, one silent and one fatal.

Silent: in a plain Markdown list, a fenced code block at column 0 closes the list item. One <ol> becomes two. Numbering still looks right because the source keeps its literal ordinals (2. renders as <ol start=2>), so nobody noticed.

Fatal: in MDX, if a JSX element spans the marker, that element is never closed and compilation fails. On temporalio/documentation this surfaces as:

Error: MDX compilation failed for .../external-storage.mdx
Cause: Expected a closing tag for `<TabItem>` (70:6-70:52) before the end of `listItem`

Any page with <Tabs> inside a numbered procedure step hits this, which currently blocks wiring those pages to snipsync at all.

Details

• Blank lines are left alone, so no trailing whitespace is introduced.
• Relative indentation inside the snippet is untouched; the existing dedent behavior is unchanged.
• Markers already at column 0 get an empty prefix, so existing targets are byte-for-byte unchanged. This is why the whole existing suite passes without fixture updates.

How this was tested

Three new tests in test/sync.test.js: indented marker inherits indentation (fence, source link, code, and relative indent all checked, plus no trailing whitespace), column-zero marker is unaffected, and re-running is idempotent.

npx jest — 38/38 pass. npx eslint . --ext .js — 0 errors, same 12 pre-existing warnings as main.

Also verified end to end against the temporalio/documentation tree with real origins:

• the previously failing page builds, and its procedure renders as a single <ol> with tabs and both code blocks intact
• the Go and Python external storage pages go from ['<ol>', '<ol start=2>'] back to ['<ol>']
• no markers leak into the rendered HTML

Docs updates needed?

No. Behavior only changes for markers that are not at column 0, which previously could not be used reliably.