fix(embed): resolve pixel tokens on export, not just off-wire

[CSSFrancis]

The bug

Under the Electron binary transport (APL_BINARY_TRANSPORT=1) a panel's pixels do not travel in the state at all: Plot2D._encode_pixels writes a "\x00bin:<adler>" change-token and the real bytes ride a PLOTBIN frame emitted by _electron._route_change.

Figure._push materialises those tokens back to inline base64 — but only if not _binary_wire(). That gate reads a process-global env var, which is on in a host app even while the push it is guarding is serialising a snapshot.

_sync_for_export() (added in 0.5.0 so exports capture widget positions where they are) re-pushes every panel from inside _repr_utils._widget_state — the one chokepoint save_html / to_html / figure_state all go through. So on a live wire, every export wrote unresolved tokens into the panel traits. A snapshot has no PLOTBIN behind it, so those are dangling references and the pixels are simply lost. Worse, because _push rewrites panel_<id>_json unconditionally, it also silently undid any materialisation the caller had performed beforehand.

The visible casualty is an add_layer overlay. _layerBytes in figure_esm.js bails on a token:

const b64 = layer.image_b64 || '';
if (!b64 || b64.charCodeAt(0) === 0) return { bytes: null, key: '' };  // token, no bytes yet

so the layer silently does not draw — while the base image still does, because imshow encodes it before the plot is attached to its Figure (no _raw_pixels side-table yet) and it is therefore plain base64, never a token. That asymmetry is what made this odd to track down in the wild: the figure looks normal, and only the overlay is missing.

Bisected to 0.5.0. Same figure, same three steps:

step	0.4.2	0.5.0 / 0.6.0 / 0.7.0
Plot2D.resolve_pixel_tokens	base64 / base64	base64 / base64
caller materialises the traits	base64 / base64	base64 / base64
export chokepoint _widget_state	base64 / base64	base64 / TOKEN

The fix

_push grows a resolve_pixels keyword; _sync_for_export passes it. An export always ships real pixels. The live wire is untouched and keeps its token/PLOTBIN split — resolving there would push megabytes of base64 through the comm on every scrub frame, which is the whole reason the transport exists.

Tests

New anyplotlib/tests/test_embed/test_export_pixel_tokens.py, 12 tests pinning both sides:

• snapshots resolve the base image, the nested layers[i].image_b64 (the copy _layerBytes actually reads), and the layer_<id>_b64 geom key; no token survives anywhere in the state
• a caller's own pre-export materialisation is no longer clobbered
• widget positions are still reconciled — the pixel fix must not cost the reason _sync_for_export exists
• the live wire still emits tokens on an ordinary push and on set_data, and returns to tokens after an export
• with no binary transport, behaviour is unchanged

6 of the 12 fail on main and all 12 pass here. Full suite green: 2007 passed, 6 skipped.