fix(parse): position-independent inline metadata parsing (#164)

Problem: `parse.body()` scanned tokens right-to-left and broke on the
first non-metadata token, so metadata only worked at the trailing end
of a line. `due:tomorrow Fix the bug` silently failed to parse the
due date.

Solution: Replace the right-to-left `while` loop with a single
left-to-right pass that extracts metadata tokens from any position.
Duplicate metadata tokens are dropped with a `log.warn`. Update docs
and tests accordingly.

doc/pending.txt

@@ -664,7 +664,7 @@ On `:w`, the description becomes `Buy milk`, the due date is stored as
 placed under the `Errands` category header.
 
 Only the first occurrence of each metadata type is consumed — duplicate
-tokens are silently dropped.
+tokens are dropped with a warning.
 
 Omnifunc completion is available for `due:`, `cat:`, and `rec:` token types.
 In insert mode, type the token prefix and press `<C-x><C-o>` to see