Restrict -flto=auto in CI to a single Linux job

Because ccache only caches object files and cannot cache linking,
enabling -flto=auto across every Release matrix job forces whole-program
LTO code generation across ~140 test and example binaries on every run
(taking 15-30 minutes even with a 100% warm ccache, and 8-12 minutes
longer on static library builds where libceres.a is re-optimized into
each test executable).

Restrict -flto=auto -Werror=odr to the ubuntu:26.04 shared-library
no-cuda job in linux.yml so ODR and LTO verification still run on every
push without paying the uncacheable LTO link cost across all other Linux,
macOS, Windows MinGW, and Android jobs.

Change-Id: I4e7f6d14b0f327a5c162fa548b3db4f77e59c344