Because it is a cache variable, it will only set the value of SOME_VAR if it is not already set. It is not wrong to do so, but usually not the best approach. I tend to only need to set plain variables in toolchain files. These may be used to initialise other cache variables (e.g. setting CMAKE_CXX_FLAGS_INIT in the toolchain file, which is used to provide the initial value for the CMAKE_CXX_FLAGS cache variable).
Yes, there is no difference between the cmake and ccmake tools in this case.
The toolchain file is pulled in whenever a new language is enabled. It is as though include(myToolchainFile.cmake) is called at the relevant line. That means any variables it sets will remain visible in that directory scope and below. So yes, values will propagate down into add_subdirectory() calls. The top level project() command is usually where languages will be enabled, but it could also be in other places like where enable_language() is called or where project() is called against with additional LANGUAGES. The file will also be pulled in for try_compile() calls, but those are in separate test sub-builds off to the side and won’t affect variables in the main project.
Probably only if you have -U options mixed in with -D options, but I can’t be sure. You can always do some experiments to find out.