Make context initialization and destruction more robust

[nnmm]

Currently, when you run e.g. ros2 run rclrs_examples minimal_publisher --ros-args -r __node:=my/node, which is an invalid node name, you'll get a message

>>> [rcutils|error_handling.c:108] rcutils_set_error_state()
This error state is being overwritten:

  'Couldn't parse remap rule: '-r __node:=my/node'. Error: Expected lexeme type 1, got 19 at index 10, at /tmp/binarydeb/ros-galactic-rcl-3.1.2/src/rcl/lexer_lookahead.c:238, at /tmp/binarydeb/ros-galactic-rcl-3.1.2/src/rcl/arguments.c:371'

with this new error message:

  'context is zero-initialized, at /tmp/binarydeb/ros-galactic-rcl-3.1.2/src/rcl/init.c:238'

rcutils_reset_error() should be called after error handling to avoid this.
<<<

This is fixed by checking that the context is indeed valid before calling rcl_shutdown() and rcl_context_fini() on it, because rcl_init() leaves the context in an invalid/zero-initialized state when it returns an error. (This is a separate issue from not resetting the error state, which we'll also do soon.)

Furthermore, the Context::new method had an unnecessary dance where the rcl_context is put in a mutex and taken out again.

I also made a tiny change so that the error code is converted to a Result right after the call, in anticipation of rcl_init_options_fini(&mut init_options).ok()? possibly overwriting the error state.