Remove nth's linear search overhead in BCF reader

[athos]

When I was profiling the BCF reader before, I found that the invocation to nth took up most of the time, which does linear search per sample.

This PR removes the overhead of the linear search and improve the performance of the BCF reader by replacing the sequential collection returned from read-typed-value with a vector, not a lazy sequence.

Here are the profiling results before and after the change:

before change	after change

By this fix, the BCF reader is now roughly 7x faster than before:

(time
 (with-open [r (vcf/reader ".cavia/large.bcf")]
   (run! (constantly nil) (vcf/read-variants-randomly r {:chr "chr1" :end 30000000} {}))))

;; before change
"Elapsed time: 7973.314958 msecs"

;; after change
"Elapsed time: 1139.505833 msecs"

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (d68c01d) 88.33% compared to head (50870bd) 88.75%.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #294      +/-   ##
==========================================
+ Coverage   88.33%   88.75%   +0.42%     
==========================================
  Files          81       83       +2     
  Lines        7028     7255     +227     
  Branches      495      515      +20     
==========================================
+ Hits         6208     6439     +231     
+ Misses        325      324       -1     
+ Partials      495      492       -3     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[matsutomo81]

Thank you for working on this!
LGTM👍

[alumi]

LGTM 👍 Thanks!