feat(hiroz): embed the bundled message definitions

[YuanYuYuan]

Closes G2 of the release-usability gaps. This now targets main. The five slices that split the original #311 are merged: #317, #318, #319, #320 and #311.

What fails without this

A downloaded hu could not decode a topic, even when a publisher was live. A release ships no message definitions, so the .msg fallback had nowhere to look.

Measured against a real talker on /chatter, with HIROZ_MSG_PATH unset:

	before	after
hu meter echo /chatter	exit 1, no .msg for std_msgs/msg/String (advertised by /chatter) was found on HIROZ_MSG_PATH	decodes: [/chatter] {"data":"Hello hiroz - #11"} — 17 messages in an 18 s window
hu meter echo on a topic nobody publishes	exit 1, names that nothing advertises a type	unchanged — still exit 1, same message

That second row is the control. Embedding must not turn a genuine failure into silence.

What this does

build.rs embeds every bundled .msg file as source text. The schema loader reads that table when HIROZ_MSG_PATH yields nothing.

• Disk still wins. A user who sets HIROZ_MSG_PATH means it. A stale embedded copy must not beat the definitions that the publisher was built from. A test pins this with a sentinel field.
• Both paths use one parser. parse_msg_string reads the embedded text and the file text, so the two cannot diverge.
• The table holds everything. The bundled set is 122 KB of text and 198 definitions. A subset would trade a fraction of a percent of binary size against "why is my type missing?".
• A test asserts that the table is non-empty and sorted. An empty table would be a silent no-op that reproduces today's behaviour.

A latent defect that embedding made reachable

hu meter pub --msg-type std_msgs/msg/String failed against a topic with a live endpoint:

topic /pub_target carries std_msgs::msg::dds_::String_, not the requested std_msgs/msg/String

The graph reports a DDS-mangled name. --msg-type takes the ROS form. The guard compared the two raw.

That guard needs a resolvable schema and a live endpoint before it runs. The one test that resolves from disk publishes to an empty topic. Nothing therefore exercised the combination until a schema could resolve without a live node.

This branch fixes it with ros_type_name_from_dds, the helper that #320 exported. A second copy of the mangling rules would drift from the first.

Breaking changes

None. This adds a fallback, and the host consults it only after the existing path fails.