Fix coverage measurement for negative line numbers

	Fixes [Bug #21220]

	Co-Authored-By: Mike Bourgeous <mike@mikebourgeous.com>
	Co-Authored-By: Jean Boussier <jean.boussier@gmail.com>

This fixes a failed assertion reported to SimpleCov

https://github.com/simplecov-ruby/simplecov/issues/1113

This can be repro'd as follows:

1. Create a file `test.rb` containing the following code

```
@foo&.(@bar)
```

2. require it with branch coverage enabled

```
ruby -rcoverage -e "Coverage.start(branches: true); require_relative 'test.rb'"
```

The assertion is failing because the Prism compiler is incorrectly
detecting the start and end cursor position of the call site for the
implicit call .()

This patch replicates the parse.y behaviour of setting the default
end_cursor to be the final closing location of the call node.

This behaviour can be verified against `parse.y` by modifying the test
command as follows:

```
ruby --parser=parse.y -rcoverage -e "Coverage.start(branches: true); require_relative 'test.rb'"
```

[Bug #20866]

Blind use of the RubyVM constant here prevents this test from
running on non-CRuby. This patch guards it with RUBY_ENGINE ==
"ruby" to make sure that doesn't happen.

Although Ruby's hashes preserve the insertion order, the ordering of
method coverage results is inherently nondeterministic, because they
are collected using `ObjectSpace`.

This reverts commit 7e0f5df2f99693267d61636d23da47f79924e9d5.

https://bugs.ruby-lang.org/issues/19857#note-7

* Add failing test.

* Add spec for eval and line coverage.

* Add test for main file coverage.