Float placement: a captured body that keeps its assignments, and floats that ride their own page

[tannevaled]

Measuring the gated single-column placer over the 157-paper arXiv corpus put it
1059 pages worse than the inline path it replaces. Two defects, both now
fixed and both read back against latex.ltx.

The body was executed without a group

All of those 1059 pages came from ONE paper. The captured float body ran with no
group of its own, so an assignment inside it escaped into the document — where
LaTeX sets a float inside \vbox\bgroup…\egroup (latex.ltx:12950), and this
engine's own inline \@float opens a \begingroup that \end@float closes.

The paper has a figure holding \put(-0.33\textwidth,0.5\textwidth){…}. \put
is undefined here, so \textwidth reads as the start of an assignment and takes
its missing number as zero — exactly what TeX itself would do. \hsize then
stayed 0pt for the remaining 230 pages, which set one word per line: 1439
pages against a reference of 333.

(Defining \put and the rest of ltpictur is worth doing on its own; it is not
what this PR is about. With the body properly grouped, an undefined command
inside a figure can no longer reach past it.)

A float could not ride the page it was written on

A float was only a candidate for a page when it was anchored BEFORE that page
began, so every float came out at least a page late and, once the text ran out,
on a float page of its own. LaTeX contributes a float at its anchor — \@xfloat
fires the output routine there with \@floatpenalty — and \@addtocurcol
(latex.ltx:15636) tests it against the room LEFT in the column: it rides the page
being built whenever \@colroom exceeds the height already set plus
\textfraction plus the float itself. That is why a figure declared halfway down
a page comes out at the top of that page.

Floats now join the page they are written on, keeping their order per caption
type — \@bitor\@currtype\@deferlist: figure 3 never precedes figure 2.

Measured, 157 arXiv papers against tectonic, GOTEX_FLOATS=1

	page error	exact	divergence (50 papers)
inline (flag off)	632	20	4.408
placer, before	1691	21	—
placer, body grouped	619	21	—
placer, both fixes	620	21	4.438

beamer is untouched: 40 pages of error over 79 decks either way. The divergence
difference is inside the noise band this measure showed when the same binary was
run over two different 50-paper samples (±0.07).

The flag stays OFF in this PR. What it changes is that turning it on is now an
improvement rather than a regression — that, with the figure geometry work of
issue #146, is the next step.

Two regression tests, each failing on main: an assignment inside a captured
figure leaves \hsize alone, and a figure written inside a page comes out on it.