fix: CSV export missing values for quaternion and RPY series

[pradhankukiran]

Fixes #1137 - quaternion and roll/pitch/yaw columns appear in CSV export with headers but no values.
Four root causes identified and fixed:

1. Off-by-one in index initialization (publisher_csv.cpp:291): indices[i] = index + 1 skipped the first in-range sample of every series. Changed to indices[i] = index, consistent with generateStatisticsCSV().

2. Machine-epsilon timestamp matching (publisher_csv.cpp:328): Values from different series were merged into a CSV row only when timestamps matched within std::numeric_limits<double>::epsilon() (~2.2e-16). Real sensor data has larger floating-point drift between channels, causing cells to go blank. Replaced with an adaptive tolerance (min_dt * 0.5) derived from the minimum sample period across all exported series — mathematically guaranteed to never merge distinct adjacent samples.

3. NaN values stalling index advancement (publisher_csv.cpp:344-348): The index only advanced when a value was not NaN, making it impossible to distinguish "no data at this timestamp" from "data is NaN at this timestamp." Added a row_used boolean vector to track participation explicitly, so indices advance correctly regardless of
   NaN status.

4. No zero-norm quaternion guard (special_messages.cpp, quaternion_to_rpy.cpp): A zero quaternion (0,0,0,0) caused 1.0 / sqrt(0) = inf, producing NaN for all RPY values, which then triggered 3rd bug. Added an early return of (0,0,0) RPY when quat_norm2 < 1e-10.

[facontidavide]

CI is red sir

[facontidavide]

nevermind, that was my fault. for the PR, actually you should return NaN if you want plotJuggler to "skip it". do not return 0.0

[pradhankukiran]

@facontidavide done, returning NaN instead of 0.0 now. thanks for finding it out.

[facontidavide]

this plugin has been retired