`naturalsize()` raises `ValueError` for custom format strings with text

Environment

• humanize 4.16.0 and current main at 42b4a1db
• Python 3.11.15

Reproduction

import humanize

print(humanize.naturalsize(999_999, gnu=True, format="Size: %.1f"))

humanize 4.15.0 returns:

Size: 976.6K

humanize 4.16.0 and current main raise:

ValueError: could not convert string to float: 'Size: 976.6'

Expected behavior

The documented custom formatter should continue to accept text around the numeric conversion.
Since this behavior worked in 4.15.0, preserving it would avoid a backward compatibility regression.

Cause

PR #329 added unit rollover handling. The new condition formats the mantissa with the
user-provided format and passes the entire result to float():

float(format % (abs_bytes / (base**exp)))

This works for a plain numeric format such as "%.1f", but not for a general formatter
such as "Size: %.1f".

Suggested fix

It might be preferable not to require the complete user-formatted result to be parseable as a float. One
backward-compatible option is to apply rollover only when the formatted result is a
plain numeric string, and otherwise preserve the pre-4.16 behavior. A more complete
solution could derive the rounded numeric mantissa independently from surrounding text.

It might also be worth adding a regression test covering both a prefix and a suffix.

How this was found

This was found by change-directed differential testing: reviewing the recent rollover change, mapping it to the public naturalsize() API, and constructing a documented custom formatter that reaches the newly added float(format % value) expression.

Related: #329, which fixes rollover but does not discuss custom format strings.

Thanks for taking the time to review this report! I appreciate all the work that goes into maintaining this project.

Beginne am naturalsize()-Einstiegspunkt und untersuche die Behandlung des Einheitenüberlaufs, die mit PR #329 eingeführt wurde, insbesondere den Pfad für benutzerdefinierte Formate, der im Report beschrieben ist. Füge eine Regressionstestabdeckung für Formatzeichenfolgen mit sowohl Präfixen als auch Suffixen hinzu und führe die relevante Testsuite aus; als erledigt gilt die Aufgabe, wenn diese Formate das Verhalten vor 4.16 beibehalten, während der Überlauf für einfache numerische Formate weiterhin funktioniert.