Ensure pprof call_tree output is a tree

When generating a call tree, pprof was using a map to keep track of all the
inline nodes for a location. That is incorrect as it may cause inline functions
at different nesting levels to reuse the same node, causing the resulting graph
to not be a tree.

When creating a tree nodes with the same info may appear on multiple places
in the tree. Keeping one of them preserves them all, which may cause disconnected
nodes to remain. To ensure the resulting graph is a connected tree, do not include
children on any removed node, which is suitable for the normal tree refinement
(nodecount and nodefraction) but does not allow visual refinement, which may eliminate
intermediate nodes. Disable visual mode refinement for call_tree to avoid this issue.

Add local symbolization using llvm-symbolizer

They were thrashing pprof's output

Add diagnostics for symbolization without mapping filenames

Add new trimproto option that generates a new profile while removing
symbol information for functions below nodefraction. This reduces the
profile sizes significantly.

Use separate mechanisms to implement graph and tree
creation, which speeds up graph creation by using a
single map from location indices to graph nodes.

pprof fetches profiles concurrently, which allows to profile multiple running
servers concurrently. However, this may translate into a large use of memory
if many profiles are merged, as pprof attempts to decode all profiles in parallel.

Limit the concurrency by chunking the concurrent fetches in groups of up to
64 profiles.

Separate implementation of graph and tree creation to speed it up.
Graph implementation maps upfront all locations to sequences of nodes,
tree implementation uses a per-parent map to keep track of a different
node per location per parent.

Non functional changes -- pre-size maps and slices to reduce
cost of building a graph out of a profile.

Profile merging spends a lot of time computing function keys, which are currently a string.
Using structs for the keys whenever possible will reduce the overhead.

The names produced by addr2line are often incomplete.  Fix this by
falling back to nm to get function names and using the nm-provided
name if it is longer.

When reconciling samples, cleanup file paths to allow functions to be
combined if their source filenames match after cleanup (eg "src.cpp" and
"./src.cpp").

Also clean up existing uses of filepath to avoid passing empty strings
to it: filepath.Base("") returns "."

internal/driver: fix spelling in help text

Add field for profile sources to indicate what is the primary
sample value, which visualizers should display by default.
Previously there was a convention for pprof to use the last
sample_index by default, this provides more flexibility.

Instead of including it the demangle package with pprof,
point to its own repo.

remove binutils wrappers used for testing

Provide native Go mocks for binutils to use during testing, instead of
relying on shell-script mocks of those tools.

Also update the build instructions to use go get