Fix g.impute() NA/NaN error with study_dates_file and check_log() IDate handling

[peterolejua]

Fixes #1508 and #1521.

g.impute()

Guards both study-date trim branches with && \!is.na(...), so a listed start or end date that is not a midnight present in the recording falls into the existing recovery branch instead of producing r4[1:(NA - 1)] = 1 / r4[NA:nrow(r4)] = 1 → NA/NaN argument.

This happens because firstmidnighti = midnightsi[grep(firstmidnight, midnights)][1] (and the matching lastmidnighti line) returns NA_integer_ (length 1) when grep() finds no match, so the existing length(...) > 0 check passes and execution reaches the subscript with NA. Two situations trigger it in practice:

• the recording starts in the evening of the listed start date, so the first midnight in the data is the next day (g.impute() errors with NA/NaN argument when study_dates_file contains a start/end date that is not a midnight in the recording (regression in 3.3-5) #1508);
• the device stopped recording before the listed end date, so that end date has no midnight in the data (discharged-device case, g.impute() errors with NA/NaN argument when study_dates_file contains a start/end date that is not a midnight in the recording (regression in 3.3-5) #1508 follow-up).

This PR also fixes a latent copy-paste in the first-midnight recovery else branch, which restored lastmidnighti/lastmidnight instead of firstmidnighti/firstmidnight.

check_log()

Adds log[] = lapply(log, as.character) at the top of check_log(), so study dates that data.table::fread() auto-parses as IDate are handled by the text-based date checks. The activity-log path is unaffected (it is already read with colClasses = "character"). Per @jhmigueles' suggestion in #1508.

Verification

• Reproduced on the original pediatric Axivity recordings from g.impute() errors with NA/NaN argument when study_dates_file contains a start/end date that is not a midnight in the recording (regression in 3.3-5) #1508 (both evening-start and discharged-device files); all process cleanly through Part 2 with the patch.
• Negative control: the same scenario on unpatched master errors with NA/NaN argument; patched runs clean.
• Added a regression test in test_StudyDatesFile.R (start/end date outside the recorded dates) and a new test_check_log.R for the IDate coercion. Both pass.