Weird num_thresholds behaviour in pr_curve plugin.

TensorBoard: v1.11.0 according to pip, also 1.11.0 when running.
TensorFlow: v1.11.0
Python: 3.7.0

The num_thresholds parameter of pr_curve_raw_data_pb from the pr_curve plugin has inconsistent and weird behaviour.

1. If num_thresholds < len(precision), then instead of approximating the PR curve in num_thresholds points, like one might expect, only the first num_thresholds points are taken. This cuts off the PR curve. There seems to be no reason to ever supply num_thresholds < len(precision), because this results in broken PR curves.

2. The solution to (1) would be passing num_thresholds=len(precision). However if pr_curve_raw_data_pb is called multiple times (i.e. to get the steps slider), it seems, that only the first num_thresholds is respected and all the consecutive PR curves are cut off.

Example:

from tensorboard.summary import pr_curve_raw_data_pb
import tensorflow as tf
import numpy as np

# [1] Curve 1 (3 data points)
pre1 = np.asarray([0, .75, 1])
rec1 = np.asarray([1, .75, 0])
tp1 = fp1 = tn1 = fn1 = np.asarray([1] * 3)

# [2] Curve 2 (5 data points)
pre2 = np.asarray([0, .7, .8, .9, 1])
rec2 = np.asarray([1, .75, .5, .25, 0])
tp2 = fp2 = tn2 = fn2 = np.asarray([1] * 5)

with tf.summary.FileWriter('./logs') as writer:
    # [3] Plotting curve 1 with proper num_thresholds = len(precision) = 3
    problem2_1 = pr_curve_raw_data_pb('problem2', tp1, fp1, tn1, fn1, pre1, rec1, num_thresholds=len(pre1))
    writer.add_summary(problem2_1, 1)

    # [4] Plotting curve 2 with the same name as [3] with proper num_thresholds = len(precision) = 5
    problem2_2 = pr_curve_raw_data_pb('problem2', tp2, fp2, tn2, fn2, pre2, rec2, num_thresholds=len(pre2))
    writer.add_summary(problem2_2, 2)

The above code results in this PR curve for step 2 (only the first 3 points are drawn).

The full curve should look like this:

A solution would be either to allow different steps of the same curve to have different num_thresholds or to pick num_thresholds evenly spread samples from precision and recall arrays.

P.S. Also it would be great, if the true_negative parameter was optional. The concept of a True Negative doesn't exist in some tasks, like detection.

Start with the pr_curve_raw_data_pb entry point and reproduce the two calls in the issue using the provided Python example. Trace how num_thresholds is handled across repeated steps and verify the behavior for arrays of different lengths. Done means consecutive curves are not unexpectedly truncated, with regression coverage for the reported cases.