Store captured samples in a hashmap during session

[osyoyu]

This patch lets Pf2 store collected samples in a hashmap backed by khashl instead of a array. This allows far lower memory consumption - the original session->samples stored every single sample collected, making the profiler's memory footprint proportional to the target program's running time, but this new implementation only records counts of distinct stacks.

Not done here:

• Make timestamp collection optional. That part is still a array
• Reconsider storage of native stacks? This patch's impl considers the tuple of (ruby_stack, pc) for dedup, but those could be managed separately. However that change would probably require changes in the serialization format.

Benchmark

Max RSS of the following script:

seconds	main	this branch	diff
18 s	382,404 KB	82,980 KB	x0.22
36 s	706,932 KB	113,024 KB	x0.16
54 s	1,039,900 KB	140,128 KB	x0.13

(script written by Codex)

# frozen_string_literal: true

require 'pf2'

DURATION = (ENV['DURATION'] || '18').to_f
INTERVAL_MS = (ENV['INTERVAL_MS'] || '1').to_i
THREADS = (ENV['THREADS'] || '4').to_i

# CPU-heavy: mix of recursion + prime checks + bit ops

def tarai(x, y, z)
  x <= y ? y : tarai(tarai(x - 1, y, z), tarai(y - 1, z, x), tarai(z - 1, x, y))
end

def is_prime(n)
  return false if n < 2
  return true if n == 2
  return false if (n & 1) == 0
  i = 3
  while i * i <= n
    return false if (n % i).zero?
    i += 2
  end
  true
end

start = Process.clock_gettime(Process::CLOCK_MONOTONIC)
profile = Pf2.profile(interval_ms: INTERVAL_MS, time_mode: :wall) do
  workers = THREADS.times.map do |idx|
    Thread.new do
      seed = 10_000 + (idx * 1000)
      while (Process.clock_gettime(Process::CLOCK_MONOTONIC) - start) < DURATION
        tarai(12, 9, 1)
        n = seed
        200.times do
          is_prime(n)
          n += 1
        end
      end
    end
  end
  workers.each(&:join)
end

STDERR.puts "profile samples=#{profile[:samples].size} functions=#{profile[:functions].size}"