Remove optionality in ActionPlan and Summary

[BillyONeal]

This is some cleanup I did while trying to better understand how the dependency planner works trying to fix microsoft/vcpkg#48797 (comment) et al.

• Fixes curl tests to work on more recent versions of curl which use 'Could not' over 'Couldn't' in more places
• Adds an 'unskipped-cascade' input to ci.ps1 tests.
• Fixes Law-of-Demeter problem in CMakeVarProvider::load_tag_vars which wanted a whole ActionPlan despite only needing the install actions
• Adds PackageSpec to ExtendedBuildResult to avoid cases where reporters were trying to re-parse the SourceControlFile et al. to figure out the package name
• Split InstallSpecSummary from SpecSummary to avoid value_or_exits when reporting about install actions. (Those values were only empty for removals which continue to use SpecSummary)
• Split InstallPlanAction into AlreadyInstalledPlanAction and InstallPlanAction, where the former has the InstalledPackageView for already installed things, and the latter has the SourceControlFileAndLocation for to-be-installed things. This avoids a lot of optionality in the rest of the product needing to deal with these two cases, particularly in parts that only ever cared about to-be-installed parts.
• Improve reporting of dependency constraint violations by reporting them as DiagnosticLine rather than LocalizedString (fixes the print color of those)
• In create_binary_control_file, loop through the feature_dependencies directly rather than trying to look each of them up by name. Note that feature_list is feature_dependencies projected to only the names.
• In commands.depend-info.cpp, don't form a vector of pointers when the real vector will do. Note that it calls create_feature_install_plan with an empty/blank status_db, so already_installed is known to be empty.
• Change perform_install_plan_action to record build time directly rather than trying to write during destructors. (This seems easier to reason about to me than the relatively non-local behavior of TrackedPackageInstallGuard, which has been deleted)

Other observations:

• ExtendedBuildResult has grown to become effectively the XxxSummary types and it is likely that it should be deleted in favor of those directly entirely.
• I was considering splitting InstallPlanAction again for the "pre-Build::compute_all_abis" and "post-Build::compute_all_abis" state. I'm not sure if that's a good change anymore though because unknown ABIs are a "normal" state of affairs, e.g. when --editable is used. I think that makes all callers of package_abi_or_empty or package_abi_or_exit suspect and those probably need to be fixed, but this PR is big enough.