docs: complete the YARD @param tags and correct six docstrings

[tas50]

Every method in lib/ already had a docstring and yard stats already reported 100.00%, but coverage counts docstrings, not tags. Measured over the .yardopts scope, three methods took a parameter that no @param described.

Before:

methods=119 missing_@param=3 missing_@return=0
PARAM:
Kitchen::Driver::Gce#state=
Kitchen::Driver::Gce#check_api_call
Kitchen::Driver::Gce#wait_for_status

After:

methods=119 missing_@param=0 missing_@return=0

The three

state= is generated by attr_accessor :state, so the reader and writer share one docstring. Putting a bare @param value on it documents the writer but makes YARD warn that #state has no such parameter, so the writer is documented separately with a @!method state=(value) directive placed after the accessor. yard is warning-free before and after.

check_api_call and wait_for_status take an explicit &block. They keep their @yield / @yieldreturn tags and gain a @param block so the documented signature matches the real one.

Six docstrings that contradicted the code

Reading each method against its docstring turned up six claims the code does not support:

Method	Claim	Reality
authorization	@return [Google::Auth::Credentials]	Google::Auth.get_application_default returns a Signet::OAuth2::Client subclass (ServiceAccountCredentials, UserRefreshCredentials, GCECredentials). Google::Auth::Credentials is an unrelated wrapper class, never in that return path.
metadata	Hash{String => String}	user_metadata only stringifies keys, so count: 3 in kitchen.yml stays an Integer. user_metadata already documented this correctly as Hash{String => Object}.
max_server_name_length	raises when a disk name "leaves no room at all"	raises once fewer than fallback_server_name_length (39) characters remain, so up to 39 characters of room can still exist.
validate!	enumerates its raise conditions	omits email, which raises under the WinRM transport.
normalize_disks	enumerates its raise conditions	omits the two it propagates from assign_boot_disk ("No disks specified", "no disk is eligible to become one"). assign_boot_disk's own docstring lists all three.
create	mutates :server_name, :hostname, :zone	also writes :created_disks (via create_instance_object → create_disks → created_disk_names) and :password for Windows guests. destroy's docstring already named :created_disks.

Prose that was already accurate is left alone.

Not changed

The YARD rake tasks (yard, yard:stats, yard:server) and .yardopts already existed; both tasks were run and are untouched.

One nit found and deliberately left, because it is a behaviour change and this PR is comments-only: normalize_disks raises "Disk name invalid. Must match #{DISK_NAME_REGEX}.", interpolating the Regexp object so the message reads (?-mix:...). The instance-name path at check_server_name uses RESOURCE_NAME_REGEX.source and reads cleanly. Worth a separate fix: PR.

Verification

• bundle exec rake test — 344 examples, 0 failures (unchanged; comments only)
• bundle exec cookstyle --chefstyle — 18 files inspected, no offenses detected (Cookstyle 9.0.0, RuboCop 1.90.0)
• bundle exec yard — no warnings
• bundle exec yard stats — 100.00% documented

The diff touches only comment lines; git diff -U0 -- lib/ filtered to non-comment lines is empty.

Merge order

No conflict with the open PRs. #176 touches spec/ only, #174 adds integration files, and #175 (README) is already merged into main. This is the sole change to lib/kitchen/driver/gce.rb, so it can merge in any order relative to those.