piecewise_linear_map: perform internal computations as f64 to increase precision

[anthrotype]

but keep the outward interface as f32 to limit the scope of changes (otherwise callers would have to deal with f64s and there are a lot)

This seems to fix #1117 in the sense that ttx_diff.py now reports Grenze-Gotisch as "output is identical".

Basically, the loss of precision happened while converting glyphs sources to ir: the instance locations are in design-space in the source; we store them as normalized location

fontc/glyphs2fontir/src/toir.rs

Line 288 in 53151a1

design_location(&axes, &m.axes_values).to_normalized(&axes_by_name),

... but then in the ir::NamedInstance they are in user-space (like in fvar) so they must be converted again to_user:

fontc/glyphs2fontir/src/source.rs

Lines 397 to 404 in 53151a1

	Some(NamedInstance {
	name: inst.name.clone(),
	location: font_info
	.locations
	.get(&inst.axes_values)
	.map(|nc| nc.to_user(&axis_map))
	.unwrap(),
	})

In this particular font, design- and user-space location are the same (there are no axis mappings). The instance at wght=200 (design) is converted to 0.666667 (normalized) and then again to 200 (user). If the lerp is done with f32 we lose precision and instead of 200 we get back 200.00002

I'm not entierly sure this is the best fix but it does seem to improve the situation at least.

[anthrotype]

no idea why I keep getting "bad credentials" while fetching GS
https://github.com/googlefonts/fontc/actions/runs/11856151327/job/33042031132?pr=1120

[cmyr]

This seems like a reasonable way to handle this, thanks for taking the ticket. 🏆